Add files using upload-large-folder tool

__init__.py

@@ -0,0 +1,4 @@
+from pipeline_kiwi_edit import KiwiEditPipeline
+from mllm_encoder import MLLMEncoder
+from conditional_embedder import ConditionalEmbedder
+from wan_video_vae import VAE

conditional_embedder.py

@@ -0,0 +1,33 @@
+import torch
+import torch.nn as nn
+from diffusers import ModelMixin, ConfigMixin
+from diffusers.configuration_utils import register_to_config
+
+
+class ConditionalEmbedder(ModelMixin, ConfigMixin):
+    """
+    Patchifies VAE-encoded conditions (source video or reference image)
+    into the DiT hidden dimension space via a Conv3d layer.
+    """
+
+    @register_to_config
+    def __init__(
+        self,
+        in_dim: int = 48,
+        dim: int = 3072,
+        patch_size: list = [1, 2, 2],
+        zero_init: bool = True,
+        ref_pad_first: bool = False,
+    ):
+        super().__init__()
+        kernel_size = tuple(patch_size)
+        self.patch_embedding = nn.Conv3d(
+            in_dim, dim, kernel_size=kernel_size, stride=kernel_size
+        )
+        self.ref_pad_first = ref_pad_first
+        if zero_init:
+            nn.init.zeros_(self.patch_embedding.weight)
+            nn.init.zeros_(self.patch_embedding.bias)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.patch_embedding(x)

pipeline_kiwi_edit.py

@@ -0,0 +1,473 @@
+import torch
+import torch.nn.functional as F
+import numpy as np
+from typing import Optional, List, Union, Callable, Tuple
+from PIL import Image, ImageOps
+from einops import rearrange
+from tqdm import tqdm
+from diffusers import DiffusionPipeline
+
+
+def sinusoidal_embedding_1d(dim, position):
+    """1D sinusoidal positional embedding for timesteps."""
+    sinusoid = torch.outer(
+        position.type(torch.float64),
+        torch.pow(
+            10000,
+            -torch.arange(dim // 2, dtype=torch.float64, device=position.device).div(
+                dim // 2
+            ),
+        ),
+    )
+    x = torch.cat([torch.cos(sinusoid), torch.sin(sinusoid)], dim=1)
+    return x.to(position.dtype)
+
+
+def _build_rope_3d(rope_module, f, h, w, device):
+    """
+    Build 3D RoPE (cos, sin) for a given (f, h, w) grid using the
+    WanRotaryPosEmbed module's precomputed buffers.
+
+    Returns:
+        (freqs_cos, freqs_sin) each of shape [1, f*h*w, 1, head_dim]
+    """
+    split_sizes = [rope_module.t_dim, rope_module.h_dim, rope_module.w_dim]
+    cos_parts = rope_module.freqs_cos.split(split_sizes, dim=1)
+    sin_parts = rope_module.freqs_sin.split(split_sizes, dim=1)
+
+    cos_f = cos_parts[0][:f].view(f, 1, 1, -1).expand(f, h, w, -1)
+    cos_h = cos_parts[1][:h].view(1, h, 1, -1).expand(f, h, w, -1)
+    cos_w = cos_parts[2][:w].view(1, 1, w, -1).expand(f, h, w, -1)
+
+    sin_f = sin_parts[0][:f].view(f, 1, 1, -1).expand(f, h, w, -1)
+    sin_h = sin_parts[1][:h].view(1, h, 1, -1).expand(f, h, w, -1)
+    sin_w = sin_parts[2][:w].view(1, 1, w, -1).expand(f, h, w, -1)
+
+    freqs_cos = torch.cat([cos_f, cos_h, cos_w], dim=-1).reshape(1, f * h * w, 1, -1).to(device)
+    freqs_sin = torch.cat([sin_f, sin_h, sin_w], dim=-1).reshape(1, f * h * w, 1, -1).to(device)
+    return freqs_cos, freqs_sin
+
+
+class KiwiEditPipeline(DiffusionPipeline):
+    """
+    Pipeline for reference-guided video and image editing using KiwiEdit.
+
+    This pipeline uses a Qwen2.5-VL multimodal LLM encoder for understanding
+    editing instructions with source visual context, a WanTransformer3DModel
+    for diffusion, and AutoencoderKLWan for VAE encoding/decoding.
+
+    Args:
+        transformer: WanTransformer3DModel - DiT backbone for denoising.
+        vae: AutoencoderKLWan - 3D causal VAE.
+        scheduler: FlowMatchEulerDiscreteScheduler or compatible scheduler.
+        mllm_encoder: MLLMEncoder - Qwen2.5-VL MLLM with learnable queries.
+        processor: AutoProcessor - Qwen2.5-VL processor/tokenizer bundle.
+        source_embedder: ConditionalEmbedder - VAE source conditioning.
+        ref_embedder: ConditionalEmbedder - VAE reference conditioning.
+    """
+
+    model_cpu_offload_seq = "mllm_encoder->source_embedder->ref_embedder->transformer->vae"
+
+    def __init__(
+        self,
+        transformer,
+        vae,
+        scheduler,
+        mllm_encoder,
+        source_embedder,
+        ref_embedder,
+        processor=None,
+    ):
+        super().__init__()
+        if isinstance(processor, (list, tuple)):
+            # Diffusers may pass the raw model_index spec; let MLLMEncoder resolve it later.
+            processor = None
+        self.register_modules(
+            transformer=transformer,
+            vae=vae,
+            scheduler=scheduler,
+            mllm_encoder=mllm_encoder,
+            processor=processor,
+            source_embedder=source_embedder,
+            ref_embedder=ref_embedder,
+        )
+        if processor is not None:
+            self.mllm_encoder.processor = processor
+
+    # ------------------------------------------------------------------ #
+    #                        Helper utilities                             #
+    # ------------------------------------------------------------------ #
+
+    @staticmethod
+    def _check_resize(height, width, num_frames, h_div=16, w_div=16, t_div=4, t_rem=1):
+        """Round height/width/num_frames to valid values."""
+        if height % h_div != 0:
+            height = (height + h_div - 1) // h_div * h_div
+        if width % w_div != 0:
+            width = (width + w_div - 1) // w_div * w_div
+        if num_frames % t_div != t_rem:
+            num_frames = (num_frames + t_div - 1) // t_div * t_div + t_rem
+        return height, width, num_frames
+
+    @staticmethod
+    def _preprocess_image(image: Image.Image, dtype, device):
+        """Convert PIL Image to tensor in [-1, 1]."""
+        arr = np.array(image, dtype=np.float32)
+        tensor = torch.from_numpy(arr).to(dtype=dtype, device=device)
+        tensor = tensor / 127.5 - 1.0  # [0, 255] -> [-1, 1]
+        tensor = tensor.permute(2, 0, 1)  # H W C -> C H W
+        return tensor
+
+    def _preprocess_video(self, frames: List[Image.Image], dtype, device):
+        """Convert list of PIL Images to tensor [1, C, T, H, W] in [-1, 1]."""
+        tensors = [self._preprocess_image(f, dtype, device) for f in frames]
+        video = torch.stack(tensors, dim=1)  # C T H W
+        return video.unsqueeze(0)  # 1 C T H W
+
+    @staticmethod
+    def _vae_output_to_video(vae_output):
+        """Convert VAE output tensor to list of PIL Images."""
+        # vae_output shape: [B, C, T, H, W] or [T, H, W, C]
+        if vae_output.dim() == 5:
+            vae_output = vae_output.squeeze(0).permute(1, 2, 3, 0)  # T H W C
+        frames = []
+        for t in range(vae_output.shape[0]):
+            frame = ((vae_output[t] + 1.0) * 127.5).clamp(0, 255)
+            frame = frame.to(device="cpu", dtype=torch.uint8).numpy()
+            frames.append(Image.fromarray(frame))
+        return frames
+
+    # ------------------------------------------------------------------ #
+    #                   Custom Flow Match Scheduler                       #
+    # ------------------------------------------------------------------ #
+
+    def _setup_scheduler(self, num_inference_steps, denoising_strength=1.0, shift=5.0):
+        """
+        Set up flow-match sigmas and timesteps matching the original diffsynth
+        FlowMatchScheduler with extra_one_step=True and shift.
+        """
+        sigma_min = 0.003 / 1.002
+        sigma_max = 1.0
+        sigma_start = sigma_min + (sigma_max - sigma_min) * denoising_strength
+        # extra_one_step: generate N+1 points, drop last
+        sigmas = torch.linspace(sigma_start, sigma_min, num_inference_steps + 1)[:-1]
+        # Apply shift
+        sigmas = shift * sigmas / (1 + (shift - 1) * sigmas)
+        timesteps = sigmas * 1000  # num_train_timesteps = 1000
+        return sigmas, timesteps
+
+    def _scheduler_step(self, model_output, sigmas, step_index, sample):
+        """Euler step for flow matching."""
+        sigma = sigmas[step_index]
+        if step_index + 1 >= len(sigmas):
+            sigma_next = 0.0
+        else:
+            sigma_next = sigmas[step_index + 1]
+        return sample + model_output * (sigma_next - sigma)
+
+    def _scheduler_add_noise(self, original_samples, noise, sigmas, step_index):
+        """Add noise at given timestep for img2img / video2video."""
+        sigma = sigmas[step_index]
+        return (1 - sigma) * original_samples + sigma * noise
+
+    def _scheduler_get_sigma(self, timestep, sigmas, timesteps):
+        """Get sigma for a given timestep."""
+        timestep_id = torch.argmin((timesteps - timestep).abs())
+        return sigmas[timestep_id]
+
+    # ------------------------------------------------------------------ #
+    #                    Transformer forward helpers                      #
+    # ------------------------------------------------------------------ #
+
+    def _model_forward(
+        self,
+        latents,
+        timestep,
+        context,
+        vae_source_input=None,
+        vae_ref_image=None,
+        sigmas=None,
+        timesteps_schedule=None,
+    ):
+        """
+        Custom DiT forward pass that handles source/ref conditioning.
+        Mirrors model_fn_wan_video from the original diffsynth pipeline.
+        """
+        device = latents.device
+        dtype = latents.dtype
+        t = self.transformer
+
+        # --- Timestep embedding ---
+        timestep_emb = sinusoidal_embedding_1d(
+            t.config.freq_dim, timestep
+        ).to(dtype)
+        time_emb = t.condition_embedder.time_embedder(timestep_emb)
+        # diffusers time_proj = Linear only (SiLU is applied separately)
+        t_mod = t.condition_embedder.time_proj(F.silu(time_emb)).unflatten(
+            1, (6, t.config.num_attention_heads * t.config.attention_head_dim)
+        )
+
+        # --- Text/context embedding ---
+        # NOTE: Do NOT apply text_embedder here. The MLLM encoder's connector
+        # already projects to dit_dim. text_embedder is for raw text encoder
+        # output (text_dim → dim), which doesn't apply to MLLM output.
+
+        # --- Patchify latents ---
+        x = latents
+        if vae_source_input is not None:
+            vae_source_cond = self.source_embedder(vae_source_input)
+            x = t.patch_embedding(x)
+            # Get sigma for this timestep
+            sigma = self._scheduler_get_sigma(timestep, sigmas, timesteps_schedule)
+            x = x + vae_source_cond * sigma
+        else:
+            x = t.patch_embedding(x)
+
+        f, h, w = x.shape[2:]
+        x = rearrange(x, "b c f h w -> b (f h w) c").contiguous()
+
+        # --- 3D RoPE frequencies (real-valued cos/sin format) ---
+        rotary_emb = _build_rope_3d(t.rope, f, h, w, device)
+
+        # --- Reference image conditioning ---
+        vae_ref_input_length = 0
+        if vae_ref_image is not None:
+            if len(vae_ref_image) > 1:
+                vae_ref = torch.cat(vae_ref_image, dim=2)  # concat along temporal
+            else:
+                vae_ref = vae_ref_image[0]
+
+            vae_ref = self.ref_embedder(vae_ref)
+            ref_f, ref_h, ref_w = vae_ref.shape[2:]
+            vae_ref = rearrange(vae_ref, "b c f h w -> b (f h w) c").contiguous()
+
+            # Recompute RoPE for extended sequence (main + ref tokens)
+            total_f = f + ref_f
+            rotary_emb = _build_rope_3d(t.rope, total_f, h, w, device)
+
+            vae_ref_input_length = vae_ref.shape[1]
+
+            if self.ref_embedder.config.ref_pad_first:
+                x = torch.cat([vae_ref, x], dim=1)
+            else:
+                x = torch.cat([x, vae_ref], dim=1)
+
+        # --- Transformer blocks ---
+        for block in t.blocks:
+            x = block(x, context, t_mod, rotary_emb)
+
+        # --- Output head ---
+        # Match diffusers' FP32 norm + modulation + projection
+        table = t.scale_shift_table
+        shift, scale = (
+            table.to(device=device) + time_emb.unsqueeze(1)
+        ).chunk(2, dim=1)
+        shift = shift.to(device=x.device)
+        scale = scale.to(device=x.device)
+        x = (t.norm_out(x.float()) * (1 + scale) + shift).type_as(x)
+        x = t.proj_out(x)
+
+        # --- Remove ref tokens from output ---
+        if vae_ref_image is not None and vae_ref_input_length > 0:
+            if self.ref_embedder.config.ref_pad_first:
+                x = x[:, vae_ref_input_length:, :]
+            else:
+                x = x[:, :-vae_ref_input_length, :]
+
+        # --- Unpatchify ---
+        patch_size = t.config.patch_size
+        x = rearrange(
+            x,
+            "b (f h w) (x y z c) -> b c (f x) (h y) (w z)",
+            f=f, h=h, w=w,
+            x=patch_size[0], y=patch_size[1], z=patch_size[2],
+        )
+        return x
+
+    # ------------------------------------------------------------------ #
+    #                          Main __call__                              #
+    # ------------------------------------------------------------------ #
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: str,
+        source_video: Optional[List[Image.Image]] = None,
+        source_input: Optional[List[Image.Image]] = None,
+        ref_image: Optional[List[Image.Image]] = None,
+        negative_prompt: Optional[str] = "",
+        input_video: Optional[List[Image.Image]] = None,
+        height: int = 480,
+        width: int = 832,
+        num_frames: int = 81,
+        num_inference_steps: int = 50,
+        guidance_scale: float = 1.0,
+        sigma_shift: float = 5.0,
+        denoising_strength: float = 1.0,
+        seed: Optional[int] = None,
+        tiled: bool = True,
+        tile_size: Tuple[int, int] = (30, 52),
+        tile_stride: Tuple[int, int] = (15, 26),
+        output_type: str = "pil",
+        progress_bar: Callable = tqdm,
+    ) -> List[Image.Image]:
+        """
+        Run KiwiEdit inference.
+
+        Args:
+            prompt: Editing instruction text.
+            source_video: Source video/image frames for MLLM context (also used as
+                source_input if source_input is not provided).
+            source_input: Source frames for VAE conditioning. If None but source_video
+                is provided, source_video is used.
+            ref_image: Optional reference image(s) for guided editing.
+            negative_prompt: Negative prompt for CFG.
+            input_video: Optional input video for video-to-video (adds noise then denoises).
+            height: Output height in pixels.
+            width: Output width in pixels.
+            num_frames: Number of output frames (1 for image editing).
+            num_inference_steps: Number of denoising steps.
+            guidance_scale: Classifier-free guidance scale.
+            sigma_shift: Flow matching shift parameter.
+            denoising_strength: How much noise to add (1.0 = full noise).
+            seed: Random seed for reproducibility.
+            tiled: Whether to use tiled VAE encoding/decoding.
+            tile_size: VAE tile size.
+            tile_stride: VAE tile stride.
+            output_type: "pil" for PIL Images, "latent" for raw latents.
+            progress_bar: Progress bar callable (e.g., tqdm).
+
+        Returns:
+            List of PIL Images (video frames).
+        """
+        device = self._execution_device
+        dtype = torch.bfloat16
+        # --- 1. Shape check ---
+        # VAE spatial factor is 16, transformer patch spatial is 2,
+        # so pixel dims must be multiples of 32.
+        height, width, num_frames = self._check_resize(
+            height, width, num_frames, h_div=32, w_div=32
+        )
+        
+        # --- 2. Determine VAE parameters ---
+        z_dim = self.vae.config.z_dim
+        # Compute upsampling factor from VAE config
+        dim_mult = self.vae.config.get("dim_mult", [1, 2, 4, 4])
+        temporal_downsample = self.vae.config.get("temperal_downsample", [False, True, True])
+        # Wan VideoVAE spatial factor is 2^(len(dim_mult)) due to extra
+        # downsampling in the encoder beyond the level transitions.
+        spatial_factor = 2 ** len(dim_mult)  # 16 for 4 levels
+        temporal_factor = 2 ** sum(temporal_downsample)  # 4 for [F, T, T]
+
+        # --- 3. MLLM encoding ---
+        context = None
+        src_video_for_mllm = source_video
+        if src_video_for_mllm is not None:
+            self.mllm_encoder._ensure_qwen_loaded()
+            if ref_image is not None:
+                # Ref mode always uses the video path (even for a single frame)
+                context = self.mllm_encoder(
+                    prompt, src_video=src_video_for_mllm, ref_image=ref_image
+                )
+            elif len(src_video_for_mllm) == 1:
+                context = self.mllm_encoder(
+                    prompt, src_image=src_video_for_mllm
+                )
+            else:
+                context = self.mllm_encoder(
+                    prompt, src_video=src_video_for_mllm
+                )
+        # For negative prompt: use zero context
+        context_nega = None
+
+        # --- 4. Setup scheduler ---
+        sigmas, timesteps = self._setup_scheduler(
+            num_inference_steps, denoising_strength, sigma_shift
+        )
+        sigmas = sigmas.to(device)
+        timesteps = timesteps.to(device)
+
+        # --- 5. Initialize noise ---
+        latent_length = (num_frames - 1) // temporal_factor + 1
+        latent_h = height // spatial_factor
+        latent_w = width // spatial_factor
+        shape = (1, z_dim, latent_length, latent_h, latent_w)
+
+        generator = None if seed is None else torch.Generator("cpu").manual_seed(seed)
+        noise = torch.randn(shape, generator=generator, device="cpu", dtype=torch.float32)
+        noise = noise.to(dtype=dtype, device=device)
+
+        # --- 6. Encode source input ---
+        vae_source_input = None
+        # Fall back to source_video if source_input not provided
+        src_for_vae = source_input if source_input is not None else source_video
+        if src_for_vae is not None:
+            src_frames = [src_for_vae[i] for i in range(min(num_frames, len(src_for_vae)))]
+            # Resize source frames to match the (possibly adjusted) target dimensions
+            src_frames = [f.resize((width, height), Image.LANCZOS) for f in src_frames]
+            src_tensor = self._preprocess_video(src_frames, dtype=torch.float32, device=device)
+            vae_source_input = self.vae.encode(src_tensor).latent_dist.sample()
+            vae_source_input = vae_source_input.to(dtype=dtype)
+
+        # --- 7. Encode reference images ---
+        vae_ref_image = None
+        if ref_image is not None:
+            vae_ref_image = []
+            for item in ref_image:
+                target_size = (width, height)
+                item = ImageOps.pad(item, target_size, color="white", centering=(0.5, 0.5))
+                ref_tensor = self._preprocess_video([item], dtype=torch.float32, device=device)
+                ref_latent = self.vae.encode(ref_tensor).latent_dist.sample()
+                vae_ref_image.append(ref_latent.to(dtype=dtype))
+
+        # --- 8. Handle input_video (video-to-video) ---
+        if input_video is not None:
+            input_tensor = self._preprocess_video(input_video, dtype=torch.float32, device=device)
+            input_latents = self.vae.encode(input_tensor).latent_dist.sample()
+            input_latents = input_latents.to(dtype=dtype)
+            latents = self._scheduler_add_noise(input_latents, noise, sigmas, 0)
+        else:
+            latents = noise
+
+        # --- 9. Denoising loop ---
+        for step_idx, timestep_val in enumerate(progress_bar(timesteps)):
+            timestep = timestep_val.unsqueeze(0).to(dtype=dtype, device=device)
+
+            # Positive prediction
+            noise_pred = self._model_forward(
+                latents=latents,
+                timestep=timestep,
+                context=context,
+                vae_source_input=vae_source_input,
+                vae_ref_image=vae_ref_image,
+                sigmas=sigmas,
+                timesteps_schedule=timesteps,
+            )
+
+            # CFG
+            # if guidance_scale != 1.0:
+            #     noise_pred_nega = self._model_forward(
+            #         latents=latents,
+            #         timestep=timestep,
+            #         context=context_nega,
+            #         vae_source_input=vae_source_input,
+            #         vae_ref_image=vae_ref_image,
+            #         sigmas=sigmas,
+            #         timesteps_schedule=timesteps,
+            #     )
+            #     noise_pred = noise_pred_nega + guidance_scale * (
+            #         noise_pred_posi - noise_pred_nega
+            #     )
+            # else:
+            #     noise_pred = noise_pred_posi
+
+            # Scheduler step
+            latents = self._scheduler_step(noise_pred, sigmas, step_idx, latents)
+
+        # --- 10. Decode ---
+        if output_type == "latent":
+            return latents
+
+        video = self.vae.decode(latents).sample
+        video = self._vae_output_to_video(video)
+        return video

ref_embedder/conditional_embedder.py

@@ -0,0 +1,33 @@
+import torch
+import torch.nn as nn
+from diffusers import ModelMixin, ConfigMixin
+from diffusers.configuration_utils import register_to_config
+
+
+class ConditionalEmbedder(ModelMixin, ConfigMixin):
+    """
+    Patchifies VAE-encoded conditions (source video or reference image)
+    into the DiT hidden dimension space via a Conv3d layer.
+    """
+
+    @register_to_config
+    def __init__(
+        self,
+        in_dim: int = 48,
+        dim: int = 3072,
+        patch_size: list = [1, 2, 2],
+        zero_init: bool = True,
+        ref_pad_first: bool = False,
+    ):
+        super().__init__()
+        kernel_size = tuple(patch_size)
+        self.patch_embedding = nn.Conv3d(
+            in_dim, dim, kernel_size=kernel_size, stride=kernel_size
+        )
+        self.ref_pad_first = ref_pad_first
+        if zero_init:
+            nn.init.zeros_(self.patch_embedding.weight)
+            nn.init.zeros_(self.patch_embedding.bias)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.patch_embedding(x)

source_embedder/conditional_embedder.py

@@ -0,0 +1,33 @@
+import torch
+import torch.nn as nn
+from diffusers import ModelMixin, ConfigMixin
+from diffusers.configuration_utils import register_to_config
+
+
+class ConditionalEmbedder(ModelMixin, ConfigMixin):
+    """
+    Patchifies VAE-encoded conditions (source video or reference image)
+    into the DiT hidden dimension space via a Conv3d layer.
+    """
+
+    @register_to_config
+    def __init__(
+        self,
+        in_dim: int = 48,
+        dim: int = 3072,
+        patch_size: list = [1, 2, 2],
+        zero_init: bool = True,
+        ref_pad_first: bool = False,
+    ):
+        super().__init__()
+        kernel_size = tuple(patch_size)
+        self.patch_embedding = nn.Conv3d(
+            in_dim, dim, kernel_size=kernel_size, stride=kernel_size
+        )
+        self.ref_pad_first = ref_pad_first
+        if zero_init:
+            nn.init.zeros_(self.patch_embedding.weight)
+            nn.init.zeros_(self.patch_embedding.bias)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.patch_embedding(x)

vae/wan_video_vae.py

@@ -0,0 +1,1486 @@
+from dataclasses import dataclass
+from typing import List, Optional
+
+from einops import rearrange, repeat
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from tqdm import tqdm
+from diffusers import ModelMixin, ConfigMixin
+from diffusers.configuration_utils import register_to_config
+
+CACHE_T = 2
+
+
+def check_is_instance(model, module_class):
+    if isinstance(model, module_class):
+        return True
+    if hasattr(model, "module") and isinstance(model.module, module_class):
+        return True
+    return False
+
+
+def block_causal_mask(x, block_size):
+    # params
+    b, n, s, _, device = *x.size(), x.device
+    assert s % block_size == 0
+    num_blocks = s // block_size
+
+    # build mask
+    mask = torch.zeros(b, n, s, s, dtype=torch.bool, device=device)
+    for i in range(num_blocks):
+        mask[:, :,
+             i * block_size:(i + 1) * block_size, :(i + 1) * block_size] = 1
+    return mask
+
+
+class CausalConv3d(nn.Conv3d):
+    """
+    Causal 3d convolusion.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._padding = (self.padding[2], self.padding[2], self.padding[1],
+                         self.padding[1], 2 * self.padding[0], 0)
+        self.padding = (0, 0, 0)
+
+    def forward(self, x, cache_x=None):
+        padding = list(self._padding)
+        if cache_x is not None and self._padding[4] > 0:
+            cache_x = cache_x.to(x.device)
+            x = torch.cat([cache_x, x], dim=2)
+            padding[4] -= cache_x.shape[2]
+        x = F.pad(x, padding)
+
+        return super().forward(x)
+
+
+class RMS_norm(nn.Module):
+
+    def __init__(self, dim, channel_first=True, images=True, bias=False):
+        super().__init__()
+        broadcastable_dims = (1, 1, 1) if not images else (1, 1)
+        shape = (dim, *broadcastable_dims) if channel_first else (dim,)
+
+        self.channel_first = channel_first
+        self.scale = dim**0.5
+        self.gamma = nn.Parameter(torch.ones(shape))
+        self.bias = nn.Parameter(torch.zeros(shape)) if bias else 0.
+
+    def forward(self, x):
+        return F.normalize(
+            x, dim=(1 if self.channel_first else
+                    -1)) * self.scale * self.gamma + self.bias
+
+
+class Upsample(nn.Upsample):
+
+    def forward(self, x):
+        """
+        Fix bfloat16 support for nearest neighbor interpolation.
+        """
+        return super().forward(x.float()).type_as(x)
+
+
+class Resample(nn.Module):
+
+    def __init__(self, dim, mode):
+        assert mode in ('none', 'upsample2d', 'upsample3d', 'downsample2d',
+                        'downsample3d')
+        super().__init__()
+        self.dim = dim
+        self.mode = mode
+
+        # layers
+        if mode == 'upsample2d':
+            self.resample = nn.Sequential(
+                Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                nn.Conv2d(dim, dim // 2, 3, padding=1))
+        elif mode == 'upsample3d':
+            self.resample = nn.Sequential(
+                Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                nn.Conv2d(dim, dim // 2, 3, padding=1))
+            self.time_conv = CausalConv3d(dim,
+                                          dim * 2, (3, 1, 1),
+                                          padding=(1, 0, 0))
+
+        elif mode == 'downsample2d':
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                nn.Conv2d(dim, dim, 3, stride=(2, 2)))
+        elif mode == 'downsample3d':
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                nn.Conv2d(dim, dim, 3, stride=(2, 2)))
+            self.time_conv = CausalConv3d(dim,
+                                          dim, (3, 1, 1),
+                                          stride=(2, 1, 1),
+                                          padding=(0, 0, 0))
+
+        else:
+            self.resample = nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        b, c, t, h, w = x.size()
+        if self.mode == 'upsample3d':
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = 'Rep'
+                    feat_idx[0] += 1
+                else:
+
+                    cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                    if cache_x.shape[2] < 2 and feat_cache[
+                            idx] is not None and feat_cache[idx] != 'Rep':
+                        # cache last frame of last two chunk
+                        cache_x = torch.cat([
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device), cache_x
+                        ],
+                                            dim=2)
+                    if cache_x.shape[2] < 2 and feat_cache[
+                            idx] is not None and feat_cache[idx] == 'Rep':
+                        cache_x = torch.cat([
+                            torch.zeros_like(cache_x).to(cache_x.device),
+                            cache_x
+                        ],
+                                            dim=2)
+                    if feat_cache[idx] == 'Rep':
+                        x = self.time_conv(x)
+                    else:
+                        x = self.time_conv(x, feat_cache[idx])
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+
+                    x = x.reshape(b, 2, c, t, h, w)
+                    x = torch.stack((x[:, 0, :, :, :, :], x[:, 1, :, :, :, :]),
+                                    3)
+                    x = x.reshape(b, c, t * 2, h, w)
+        t = x.shape[2]
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+        x = self.resample(x)
+        x = rearrange(x, '(b t) c h w -> b c t h w', t=t)
+
+        if self.mode == 'downsample3d':
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = x.clone()
+                    feat_idx[0] += 1
+                else:
+                    cache_x = x[:, :, -1:, :, :].clone()
+                    x = self.time_conv(
+                        torch.cat([feat_cache[idx][:, :, -1:, :, :], x], 2))
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+        return x
+
+    def init_weight(self, conv):
+        conv_weight = conv.weight
+        nn.init.zeros_(conv_weight)
+        c1, c2, t, h, w = conv_weight.size()
+        one_matrix = torch.eye(c1, c2)
+        init_matrix = one_matrix
+        nn.init.zeros_(conv_weight)
+        conv_weight.data[:, :, 1, 0, 0] = init_matrix
+        conv.weight.data.copy_(conv_weight)
+        nn.init.zeros_(conv.bias.data)
+
+    def init_weight2(self, conv):
+        conv_weight = conv.weight.data
+        nn.init.zeros_(conv_weight)
+        c1, c2, t, h, w = conv_weight.size()
+        init_matrix = torch.eye(c1 // 2, c2)
+        conv_weight[:c1 // 2, :, -1, 0, 0] = init_matrix
+        conv_weight[c1 // 2:, :, -1, 0, 0] = init_matrix
+        conv.weight.data.copy_(conv_weight)
+        nn.init.zeros_(conv.bias.data)
+
+
+
+def patchify(x, patch_size):
+    if patch_size == 1:
+        return x
+    if x.dim() == 4:
+        x = rearrange(x, "b c (h q) (w r) -> b (c r q) h w", q=patch_size, r=patch_size)
+    elif x.dim() == 5:
+        x = rearrange(x,
+                      "b c f (h q) (w r) -> b (c r q) f h w",
+                      q=patch_size,
+                      r=patch_size)
+    else:
+        raise ValueError(f"Invalid input shape: {x.shape}")
+    return x
+
+
+def unpatchify(x, patch_size):
+    if patch_size == 1:
+        return x
+    if x.dim() == 4:
+        x = rearrange(x, "b (c r q) h w -> b c (h q) (w r)", q=patch_size, r=patch_size)
+    elif x.dim() == 5:
+        x = rearrange(x,
+                      "b (c r q) f h w -> b c f (h q) (w r)",
+                      q=patch_size,
+                      r=patch_size)
+    return x
+
+
+class Resample38(Resample):
+
+    def __init__(self, dim, mode):
+        assert mode in (
+            "none",
+            "upsample2d",
+            "upsample3d",
+            "downsample2d",
+            "downsample3d",
+        )
+        super(Resample, self).__init__()
+        self.dim = dim
+        self.mode = mode
+
+        # layers
+        if mode == "upsample2d":
+            self.resample = nn.Sequential(
+                Upsample(scale_factor=(2.0, 2.0), mode="nearest-exact"),
+                nn.Conv2d(dim, dim, 3, padding=1),
+            )
+        elif mode == "upsample3d":
+            self.resample = nn.Sequential(
+                Upsample(scale_factor=(2.0, 2.0), mode="nearest-exact"),
+                nn.Conv2d(dim, dim, 3, padding=1),
+            )
+            self.time_conv = CausalConv3d(dim, dim * 2, (3, 1, 1), padding=(1, 0, 0))
+        elif mode == "downsample2d":
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)), nn.Conv2d(dim, dim, 3, stride=(2, 2))
+            )
+        elif mode == "downsample3d":
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)), nn.Conv2d(dim, dim, 3, stride=(2, 2))
+            )
+            self.time_conv = CausalConv3d(
+                dim, dim, (3, 1, 1), stride=(2, 1, 1), padding=(0, 0, 0)
+            )
+        else:
+            self.resample = nn.Identity()
+
+class ResidualBlock(nn.Module):
+
+    def __init__(self, in_dim, out_dim, dropout=0.0):
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+
+        # layers
+        self.residual = nn.Sequential(
+            RMS_norm(in_dim, images=False), nn.SiLU(),
+            CausalConv3d(in_dim, out_dim, 3, padding=1),
+            RMS_norm(out_dim, images=False), nn.SiLU(), nn.Dropout(dropout),
+            CausalConv3d(out_dim, out_dim, 3, padding=1))
+        self.shortcut = CausalConv3d(in_dim, out_dim, 1) \
+            if in_dim != out_dim else nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        h = self.shortcut(x)
+        for layer in self.residual:
+            if check_is_instance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x + h
+
+
+class AttentionBlock(nn.Module):
+    """
+    Causal self-attention with a single head.
+    """
+
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+
+        # layers
+        self.norm = RMS_norm(dim)
+        self.to_qkv = nn.Conv2d(dim, dim * 3, 1)
+        self.proj = nn.Conv2d(dim, dim, 1)
+
+        # zero out the last layer params
+        nn.init.zeros_(self.proj.weight)
+
+    def forward(self, x):
+        identity = x
+        b, c, t, h, w = x.size()
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+        x = self.norm(x)
+        # compute query, key, value
+        q, k, v = self.to_qkv(x).reshape(b * t, 1, c * 3, -1).permute(
+            0, 1, 3, 2).contiguous().chunk(3, dim=-1)
+
+        # apply attention
+        x = F.scaled_dot_product_attention(
+            q,
+            k,
+            v,
+            #attn_mask=block_causal_mask(q, block_size=h * w)
+        )
+        x = x.squeeze(1).permute(0, 2, 1).reshape(b * t, c, h, w)
+
+        # output
+        x = self.proj(x)
+        x = rearrange(x, '(b t) c h w-> b c t h w', t=t)
+        return x + identity
+
+
+class AvgDown3D(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        factor_t,
+        factor_s=1,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.factor_t = factor_t
+        self.factor_s = factor_s
+        self.factor = self.factor_t * self.factor_s * self.factor_s
+
+        assert in_channels * self.factor % out_channels == 0
+        self.group_size = in_channels * self.factor // out_channels
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        pad_t = (self.factor_t - x.shape[2] % self.factor_t) % self.factor_t
+        pad = (0, 0, 0, 0, pad_t, 0)
+        x = F.pad(x, pad)
+        B, C, T, H, W = x.shape
+        x = x.view(
+            B,
+            C,
+            T // self.factor_t,
+            self.factor_t,
+            H // self.factor_s,
+            self.factor_s,
+            W // self.factor_s,
+            self.factor_s,
+        )
+        x = x.permute(0, 1, 3, 5, 7, 2, 4, 6).contiguous()
+        x = x.view(
+            B,
+            C * self.factor,
+            T // self.factor_t,
+            H // self.factor_s,
+            W // self.factor_s,
+        )
+        x = x.view(
+            B,
+            self.out_channels,
+            self.group_size,
+            T // self.factor_t,
+            H // self.factor_s,
+            W // self.factor_s,
+        )
+        x = x.mean(dim=2)
+        return x
+
+
+class DupUp3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        factor_t,
+        factor_s=1,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        self.factor_t = factor_t
+        self.factor_s = factor_s
+        self.factor = self.factor_t * self.factor_s * self.factor_s
+
+        assert out_channels * self.factor % in_channels == 0
+        self.repeats = out_channels * self.factor // in_channels
+
+    def forward(self, x: torch.Tensor, first_chunk=False) -> torch.Tensor:
+        x = x.repeat_interleave(self.repeats, dim=1)
+        x = x.view(
+            x.size(0),
+            self.out_channels,
+            self.factor_t,
+            self.factor_s,
+            self.factor_s,
+            x.size(2),
+            x.size(3),
+            x.size(4),
+        )
+        x = x.permute(0, 1, 5, 2, 6, 3, 7, 4).contiguous()
+        x = x.view(
+            x.size(0),
+            self.out_channels,
+            x.size(2) * self.factor_t,
+            x.size(4) * self.factor_s,
+            x.size(6) * self.factor_s,
+        )
+        if first_chunk:
+            x = x[:, :, self.factor_t - 1 :, :, :]
+        return x
+
+
+class Down_ResidualBlock(nn.Module):
+    def __init__(
+        self, in_dim, out_dim, dropout, mult, temperal_downsample=False, down_flag=False
+    ):
+        super().__init__()
+
+        # Shortcut path with downsample
+        self.avg_shortcut = AvgDown3D(
+            in_dim,
+            out_dim,
+            factor_t=2 if temperal_downsample else 1,
+            factor_s=2 if down_flag else 1,
+        )
+
+        # Main path with residual blocks and downsample
+        downsamples = []
+        for _ in range(mult):
+            downsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+            in_dim = out_dim
+
+        # Add the final downsample block
+        if down_flag:
+            mode = "downsample3d" if temperal_downsample else "downsample2d"
+            downsamples.append(Resample38(out_dim, mode=mode))
+
+        self.downsamples = nn.Sequential(*downsamples)
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        x_copy = x.clone()
+        for module in self.downsamples:
+            x = module(x, feat_cache, feat_idx)
+
+        return x + self.avg_shortcut(x_copy)
+
+
+class Up_ResidualBlock(nn.Module):
+    def __init__(
+        self, in_dim, out_dim, dropout, mult, temperal_upsample=False, up_flag=False
+    ):
+        super().__init__()
+        # Shortcut path with upsample
+        if up_flag:
+            self.avg_shortcut = DupUp3D(
+                in_dim,
+                out_dim,
+                factor_t=2 if temperal_upsample else 1,
+                factor_s=2 if up_flag else 1,
+            )
+        else:
+            self.avg_shortcut = None
+
+        # Main path with residual blocks and upsample
+        upsamples = []
+        for _ in range(mult):
+            upsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+            in_dim = out_dim
+
+        # Add the final upsample block
+        if up_flag:
+            mode = "upsample3d" if temperal_upsample else "upsample2d"
+            upsamples.append(Resample38(out_dim, mode=mode))
+
+        self.upsamples = nn.Sequential(*upsamples)
+
+    def forward(self, x, feat_cache=None, feat_idx=[0], first_chunk=False):
+        x_main = x.clone()
+        for module in self.upsamples:
+            x_main = module(x_main, feat_cache, feat_idx)
+        if self.avg_shortcut is not None:
+            x_shortcut = self.avg_shortcut(x, first_chunk)
+            return x_main + x_shortcut
+        else:
+            return x_main
+
+
+class Encoder3d(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[True, True, False],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+
+        # dimensions
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        # init block
+        self.conv1 = CausalConv3d(3, dims[0], 3, padding=1)
+
+        # downsample blocks
+        downsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            for _ in range(num_res_blocks):
+                downsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    downsamples.append(AttentionBlock(out_dim))
+                in_dim = out_dim
+
+            # downsample block
+            if i != len(dim_mult) - 1:
+                mode = 'downsample3d' if temperal_downsample[
+                    i] else 'downsample2d'
+                downsamples.append(Resample(out_dim, mode=mode))
+                scale /= 2.0
+        self.downsamples = nn.Sequential(*downsamples)
+
+        # middle blocks
+        self.middle = nn.Sequential(ResidualBlock(out_dim, out_dim, dropout),
+                                    AttentionBlock(out_dim),
+                                    ResidualBlock(out_dim, out_dim, dropout))
+
+        # output blocks
+        self.head = nn.Sequential(RMS_norm(out_dim, images=False), nn.SiLU(),
+                                  CausalConv3d(out_dim, z_dim, 3, padding=1))
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([
+                    feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                        cache_x.device), cache_x
+                ],
+                                    dim=2)
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## downsamples
+        for layer in self.downsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## middle
+        for layer in self.middle:
+            if check_is_instance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if check_is_instance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+class Encoder3d_38(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[False, True, True],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+
+        # dimensions
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        # init block
+        self.conv1 = CausalConv3d(12, dims[0], 3, padding=1)
+
+        # downsample blocks
+        downsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            t_down_flag = (
+                temperal_downsample[i] if i < len(temperal_downsample) else False
+            )
+            downsamples.append(
+                Down_ResidualBlock(
+                    in_dim=in_dim,
+                    out_dim=out_dim,
+                    dropout=dropout,
+                    mult=num_res_blocks,
+                    temperal_downsample=t_down_flag,
+                    down_flag=i != len(dim_mult) - 1,
+                )
+            )
+            scale /= 2.0
+        self.downsamples = nn.Sequential(*downsamples)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(out_dim, out_dim, dropout),
+            AttentionBlock(out_dim),
+            ResidualBlock(out_dim, out_dim, dropout),
+        )
+
+        # # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False),
+            nn.SiLU(),
+            CausalConv3d(out_dim, z_dim, 3, padding=1),
+        )
+
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device),
+                        cache_x,
+                    ],
+                    dim=2,
+                )
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## downsamples
+        for layer in self.downsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## middle
+        for layer in self.middle:
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    cache_x = torch.cat(
+                        [
+                            feat_cache[idx][:, :, -1, :, :]
+                            .unsqueeze(2)
+                            .to(cache_x.device),
+                            cache_x,
+                        ],
+                        dim=2,
+                    )
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+
+        return x
+
+
+class Decoder3d(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_upsample=[False, True, True],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_upsample = temperal_upsample
+
+        # dimensions
+        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
+        scale = 1.0 / 2**(len(dim_mult) - 2)
+
+        # init block
+        self.conv1 = CausalConv3d(z_dim, dims[0], 3, padding=1)
+
+        # middle blocks
+        self.middle = nn.Sequential(ResidualBlock(dims[0], dims[0], dropout),
+                                    AttentionBlock(dims[0]),
+                                    ResidualBlock(dims[0], dims[0], dropout))
+
+        # upsample blocks
+        upsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            if i == 1 or i == 2 or i == 3:
+                in_dim = in_dim // 2
+            for _ in range(num_res_blocks + 1):
+                upsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    upsamples.append(AttentionBlock(out_dim))
+                in_dim = out_dim
+
+            # upsample block
+            if i != len(dim_mult) - 1:
+                mode = 'upsample3d' if temperal_upsample[i] else 'upsample2d'
+                upsamples.append(Resample(out_dim, mode=mode))
+                scale *= 2.0
+        self.upsamples = nn.Sequential(*upsamples)
+
+        # output blocks
+        self.head = nn.Sequential(RMS_norm(out_dim, images=False), nn.SiLU(),
+                                  CausalConv3d(out_dim, 3, 3, padding=1))
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        ## conv1
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([
+                    feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                        cache_x.device), cache_x
+                ],
+                                    dim=2)
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## middle
+        for layer in self.middle:
+            if check_is_instance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## upsamples
+        for layer in self.upsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if check_is_instance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+
+class Decoder3d_38(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_upsample=[False, True, True],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_upsample = temperal_upsample
+
+        # dimensions
+        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
+        scale = 1.0 / 2 ** (len(dim_mult) - 2)
+        # init block
+        self.conv1 = CausalConv3d(z_dim, dims[0], 3, padding=1)
+
+        # middle blocks
+        self.middle = nn.Sequential(ResidualBlock(dims[0], dims[0], dropout),
+                                    AttentionBlock(dims[0]),
+                                    ResidualBlock(dims[0], dims[0], dropout))
+
+        # upsample blocks
+        upsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            t_up_flag = temperal_upsample[i] if i < len(temperal_upsample) else False
+            upsamples.append(
+                Up_ResidualBlock(in_dim=in_dim,
+                                 out_dim=out_dim,
+                                 dropout=dropout,
+                                 mult=num_res_blocks + 1,
+                                 temperal_upsample=t_up_flag,
+                                 up_flag=i != len(dim_mult) - 1))
+        self.upsamples = nn.Sequential(*upsamples)
+
+        # output blocks
+        self.head = nn.Sequential(RMS_norm(out_dim, images=False), nn.SiLU(),
+                                  CausalConv3d(out_dim, 12, 3, padding=1))
+
+
+    def forward(self, x, feat_cache=None, feat_idx=[0], first_chunk=False):
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device),
+                        cache_x,
+                    ],
+                    dim=2,
+                )
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        for layer in self.middle:
+            if check_is_instance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## upsamples
+        for layer in self.upsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx, first_chunk)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if check_is_instance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    cache_x = torch.cat(
+                        [
+                            feat_cache[idx][:, :, -1, :, :]
+                            .unsqueeze(2)
+                            .to(cache_x.device),
+                            cache_x,
+                        ],
+                        dim=2,
+                    )
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+def count_conv3d(model):
+    count = 0
+    for m in model.modules():
+        if isinstance(m, CausalConv3d):
+            count += 1
+    return count
+
+
+class VideoVAE_(nn.Module):
+
+    def __init__(self,
+                 dim=96,
+                 z_dim=16,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[False, True, True],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+        self.temperal_upsample = temperal_downsample[::-1]
+
+        # modules
+        self.encoder = Encoder3d(dim, z_dim * 2, dim_mult, num_res_blocks,
+                                 attn_scales, self.temperal_downsample, dropout)
+        self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
+        self.conv2 = CausalConv3d(z_dim, z_dim, 1)
+        self.decoder = Decoder3d(dim, z_dim, dim_mult, num_res_blocks,
+                                 attn_scales, self.temperal_upsample, dropout)
+
+    def forward(self, x):
+        mu, log_var = self.encode(x)
+        z = self.reparameterize(mu, log_var)
+        x_recon = self.decode(z)
+        return x_recon, mu, log_var
+
+    def encode(self, x, scale):
+        self.clear_cache()
+        ## cache
+        t = x.shape[2]
+        iter_ = 1 + (t - 1) // 4
+
+        for i in range(iter_):
+            self._enc_conv_idx = [0]
+            if i == 0:
+                out = self.encoder(x[:, :, :1, :, :],
+                                   feat_cache=self._enc_feat_map,
+                                   feat_idx=self._enc_conv_idx)
+            else:
+                out_ = self.encoder(x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
+                                    feat_cache=self._enc_feat_map,
+                                    feat_idx=self._enc_conv_idx)
+                out = torch.cat([out, out_], 2)
+        mu, log_var = self.conv1(out).chunk(2, dim=1)
+        if isinstance(scale[0], torch.Tensor):
+            scale = [s.to(dtype=mu.dtype, device=mu.device) for s in scale]
+            mu = (mu - scale[0].view(1, self.z_dim, 1, 1, 1)) * scale[1].view(
+                1, self.z_dim, 1, 1, 1)
+        else:
+            scale = scale.to(dtype=mu.dtype, device=mu.device)
+            mu = (mu - scale[0]) * scale[1]
+        return mu
+
+    def decode(self, z, scale):
+        self.clear_cache()
+        # z: [b,c,t,h,w]
+        if isinstance(scale[0], torch.Tensor):
+            scale = [s.to(dtype=z.dtype, device=z.device) for s in scale]
+            z = z / scale[1].view(1, self.z_dim, 1, 1, 1) + scale[0].view(
+                1, self.z_dim, 1, 1, 1)
+        else:
+            scale = scale.to(dtype=z.dtype, device=z.device)
+            z = z / scale[1] + scale[0]
+        iter_ = z.shape[2]
+        x = self.conv2(z)
+        for i in range(iter_):
+            self._conv_idx = [0]
+            if i == 0:
+                out = self.decoder(x[:, :, i:i + 1, :, :],
+                                   feat_cache=self._feat_map,
+                                   feat_idx=self._conv_idx)
+            else:
+                out_ = self.decoder(x[:, :, i:i + 1, :, :],
+                                    feat_cache=self._feat_map,
+                                    feat_idx=self._conv_idx)
+                out = torch.cat([out, out_], 2) # may add tensor offload
+        return out
+
+    def reparameterize(self, mu, log_var):
+        std = torch.exp(0.5 * log_var)
+        eps = torch.randn_like(std)
+        return eps * std + mu
+
+    def sample(self, imgs, deterministic=False):
+        mu, log_var = self.encode(imgs)
+        if deterministic:
+            return mu
+        std = torch.exp(0.5 * log_var.clamp(-30.0, 20.0))
+        return mu + std * torch.randn_like(std)
+
+    def clear_cache(self):
+        self._conv_num = count_conv3d(self.decoder)
+        self._conv_idx = [0]
+        self._feat_map = [None] * self._conv_num
+        # cache encode
+        self._enc_conv_num = count_conv3d(self.encoder)
+        self._enc_conv_idx = [0]
+        self._enc_feat_map = [None] * self._enc_conv_num
+
+
+class WanVideoVAE(nn.Module):
+
+    def __init__(self, z_dim=16):
+        super().__init__()
+
+        mean = [
+            -0.7571, -0.7089, -0.9113, 0.1075, -0.1745, 0.9653, -0.1517, 1.5508,
+            0.4134, -0.0715, 0.5517, -0.3632, -0.1922, -0.9497, 0.2503, -0.2921
+        ]
+        std = [
+            2.8184, 1.4541, 2.3275, 2.6558, 1.2196, 1.7708, 2.6052, 2.0743,
+            3.2687, 2.1526, 2.8652, 1.5579, 1.6382, 1.1253, 2.8251, 1.9160
+        ]
+        self.mean = torch.tensor(mean)
+        self.std = torch.tensor(std)
+        self.scale = [self.mean, 1.0 / self.std]
+
+        # init model
+        self.model = VideoVAE_(z_dim=z_dim).eval().requires_grad_(False)
+        self.upsampling_factor = 8
+        self.z_dim = z_dim
+
+
+    def build_1d_mask(self, length, left_bound, right_bound, border_width):
+        x = torch.ones((length,))
+        if not left_bound:
+            x[:border_width] = (torch.arange(border_width) + 1) / border_width
+        if not right_bound:
+            x[-border_width:] = torch.flip((torch.arange(border_width) + 1) / border_width, dims=(0,))
+        return x
+
+
+    def build_mask(self, data, is_bound, border_width):
+        _, _, _, H, W = data.shape
+        h = self.build_1d_mask(H, is_bound[0], is_bound[1], border_width[0])
+        w = self.build_1d_mask(W, is_bound[2], is_bound[3], border_width[1])
+
+        h = repeat(h, "H -> H W", H=H, W=W)
+        w = repeat(w, "W -> H W", H=H, W=W)
+
+        mask = torch.stack([h, w]).min(dim=0).values
+        mask = rearrange(mask, "H W -> 1 1 1 H W")
+        return mask
+
+
+    def tiled_decode(self, hidden_states, device, tile_size, tile_stride):
+        _, _, T, H, W = hidden_states.shape
+        size_h, size_w = tile_size
+        stride_h, stride_w = tile_stride
+
+        # Split tasks
+        tasks = []
+        for h in range(0, H, stride_h):
+            if (h-stride_h >= 0 and h-stride_h+size_h >= H): continue
+            for w in range(0, W, stride_w):
+                if (w-stride_w >= 0 and w-stride_w+size_w >= W): continue
+                h_, w_ = h + size_h, w + size_w
+                tasks.append((h, h_, w, w_))
+
+        data_device = "cpu"
+        computation_device = device
+
+        out_T = T * 4 - 3
+        weight = torch.zeros((1, 1, out_T, H * self.upsampling_factor, W * self.upsampling_factor), dtype=hidden_states.dtype, device=data_device)
+        values = torch.zeros((1, 3, out_T, H * self.upsampling_factor, W * self.upsampling_factor), dtype=hidden_states.dtype, device=data_device)
+
+        for h, h_, w, w_ in tqdm(tasks, desc="VAE decoding"):
+            hidden_states_batch = hidden_states[:, :, :, h:h_, w:w_].to(computation_device)
+            hidden_states_batch = self.model.decode(hidden_states_batch, self.scale).to(data_device)
+
+            mask = self.build_mask(
+                hidden_states_batch,
+                is_bound=(h==0, h_>=H, w==0, w_>=W),
+                border_width=((size_h - stride_h) * self.upsampling_factor, (size_w - stride_w) * self.upsampling_factor)
+            ).to(dtype=hidden_states.dtype, device=data_device)
+
+            target_h = h * self.upsampling_factor
+            target_w = w * self.upsampling_factor
+            values[
+                :,
+                :,
+                :,
+                target_h:target_h + hidden_states_batch.shape[3],
+                target_w:target_w + hidden_states_batch.shape[4],
+            ] += hidden_states_batch * mask
+            weight[
+                :,
+                :,
+                :,
+                target_h: target_h + hidden_states_batch.shape[3],
+                target_w: target_w + hidden_states_batch.shape[4],
+            ] += mask
+        values = values / weight
+        values = values.clamp_(-1, 1)
+        return values
+
+
+    def tiled_encode(self, video, device, tile_size, tile_stride):
+        _, _, T, H, W = video.shape
+        size_h, size_w = tile_size
+        stride_h, stride_w = tile_stride
+
+        # Split tasks
+        tasks = []
+        for h in range(0, H, stride_h):
+            if (h-stride_h >= 0 and h-stride_h+size_h >= H): continue
+            for w in range(0, W, stride_w):
+                if (w-stride_w >= 0 and w-stride_w+size_w >= W): continue
+                h_, w_ = h + size_h, w + size_w
+                tasks.append((h, h_, w, w_))
+
+        data_device = "cpu"
+        computation_device = device
+
+        out_T = (T + 3) // 4
+        weight = torch.zeros((1, 1, out_T, H // self.upsampling_factor, W // self.upsampling_factor), dtype=video.dtype, device=data_device)
+        values = torch.zeros((1, self.z_dim, out_T, H // self.upsampling_factor, W // self.upsampling_factor), dtype=video.dtype, device=data_device)
+
+        for h, h_, w, w_ in tqdm(tasks, desc="VAE encoding"):
+            hidden_states_batch = video[:, :, :, h:h_, w:w_].to(computation_device)
+            hidden_states_batch = self.model.encode(hidden_states_batch, self.scale).to(data_device)
+
+            mask = self.build_mask(
+                hidden_states_batch,
+                is_bound=(h==0, h_>=H, w==0, w_>=W),
+                border_width=((size_h - stride_h) // self.upsampling_factor, (size_w - stride_w) // self.upsampling_factor)
+            ).to(dtype=video.dtype, device=data_device)
+
+            target_h = h // self.upsampling_factor
+            target_w = w // self.upsampling_factor
+            values[
+                :,
+                :,
+                :,
+                target_h:target_h + hidden_states_batch.shape[3],
+                target_w:target_w + hidden_states_batch.shape[4],
+            ] += hidden_states_batch * mask
+            weight[
+                :,
+                :,
+                :,
+                target_h: target_h + hidden_states_batch.shape[3],
+                target_w: target_w + hidden_states_batch.shape[4],
+            ] += mask
+        values = values / weight
+        return values
+
+
+    def single_encode(self, video, device):
+        video = video.to(device)
+        x = self.model.encode(video, self.scale)
+        return x
+
+
+    def single_decode(self, hidden_state, device):
+        hidden_state = hidden_state.to(device)
+        video = self.model.decode(hidden_state, self.scale)
+        return video.clamp_(-1, 1)
+
+
+    def encode(self, videos, device, tiled=False, tile_size=(34, 34), tile_stride=(18, 16)):
+        videos = [video.to("cpu") for video in videos]
+        hidden_states = []
+        for video in videos:
+            video = video.unsqueeze(0)
+            if tiled:
+                tile_size = (tile_size[0] * self.upsampling_factor, tile_size[1] * self.upsampling_factor)
+                tile_stride = (tile_stride[0] * self.upsampling_factor, tile_stride[1] * self.upsampling_factor)
+                hidden_state = self.tiled_encode(video, device, tile_size, tile_stride)
+            else:
+                hidden_state = self.single_encode(video, device)
+            hidden_state = hidden_state.squeeze(0)
+            hidden_states.append(hidden_state)
+        hidden_states = torch.stack(hidden_states)
+        return hidden_states
+
+
+    def decode(self, hidden_states, device, tiled=False, tile_size=(34, 34), tile_stride=(18, 16)):
+        hidden_states = [hidden_state.to("cpu") for hidden_state in hidden_states]
+        videos = []
+        for hidden_state in hidden_states:
+            hidden_state = hidden_state.unsqueeze(0)
+            if tiled:
+                video = self.tiled_decode(hidden_state, device, tile_size, tile_stride)
+            else:
+                video = self.single_decode(hidden_state, device)
+            video = video.squeeze(0)
+            videos.append(video)
+        videos = torch.stack(videos)
+        return videos
+
+
+    @staticmethod
+    def state_dict_converter():
+        return WanVideoVAEStateDictConverter()
+
+
+class WanVideoVAEStateDictConverter:
+
+    def __init__(self):
+        pass
+
+    def from_civitai(self, state_dict):
+        state_dict_ = {}
+        if 'model_state' in state_dict:
+            state_dict = state_dict['model_state']
+        for name in state_dict:
+            state_dict_['model.' + name] = state_dict[name]
+        return state_dict_
+
+
+class VideoVAE38_(VideoVAE_):
+
+    def __init__(self,
+                 dim=160,
+                 z_dim=48,
+                 dec_dim=256,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[False, True, True],
+                 dropout=0.0):
+        super(VideoVAE_, self).__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+        self.temperal_upsample = temperal_downsample[::-1]
+
+        # modules
+        self.encoder = Encoder3d_38(dim, z_dim * 2, dim_mult, num_res_blocks,
+                                    attn_scales, self.temperal_downsample, dropout)
+        self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
+        self.conv2 = CausalConv3d(z_dim, z_dim, 1)
+        self.decoder = Decoder3d_38(dec_dim, z_dim, dim_mult, num_res_blocks,
+                                    attn_scales, self.temperal_upsample, dropout)
+
+
+    def encode(self, x, scale):
+        self.clear_cache()
+        x = patchify(x, patch_size=2)
+        t = x.shape[2]
+        iter_ = 1 + (t - 1) // 4
+        for i in range(iter_):
+            self._enc_conv_idx = [0]
+            if i == 0:
+                out = self.encoder(x[:, :, :1, :, :],
+                                   feat_cache=self._enc_feat_map,
+                                   feat_idx=self._enc_conv_idx)
+            else:
+                out_ = self.encoder(x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
+                                    feat_cache=self._enc_feat_map,
+                                    feat_idx=self._enc_conv_idx)
+                out = torch.cat([out, out_], 2)
+        mu, log_var = self.conv1(out).chunk(2, dim=1)
+        if isinstance(scale[0], torch.Tensor):
+            scale = [s.to(dtype=mu.dtype, device=mu.device) for s in scale]
+            mu = (mu - scale[0].view(1, self.z_dim, 1, 1, 1)) * scale[1].view(
+                1, self.z_dim, 1, 1, 1)
+        else:
+            scale = scale.to(dtype=mu.dtype, device=mu.device)
+            mu = (mu - scale[0]) * scale[1]
+        self.clear_cache()
+        return mu
+
+
+    def decode(self, z, scale):
+        self.clear_cache()
+        if isinstance(scale[0], torch.Tensor):
+            scale = [s.to(dtype=z.dtype, device=z.device) for s in scale]
+            z = z / scale[1].view(1, self.z_dim, 1, 1, 1) + scale[0].view(
+                1, self.z_dim, 1, 1, 1)
+        else:
+            scale = scale.to(dtype=z.dtype, device=z.device)
+            z = z / scale[1] + scale[0]
+        iter_ = z.shape[2]
+        x = self.conv2(z)
+        for i in range(iter_):
+            self._conv_idx = [0]
+            if i == 0:
+                out = self.decoder(x[:, :, i:i + 1, :, :],
+                                   feat_cache=self._feat_map,
+                                   feat_idx=self._conv_idx,
+                                   first_chunk=True)
+            else:
+                out_ = self.decoder(x[:, :, i:i + 1, :, :],
+                                    feat_cache=self._feat_map,
+                                    feat_idx=self._conv_idx)
+                out = torch.cat([out, out_], 2)
+        out = unpatchify(out, patch_size=2)
+        self.clear_cache()
+        return out
+
+
+class WanVideoVAE38(WanVideoVAE):
+
+    def __init__(self, z_dim=48, dim=160):
+        super(WanVideoVAE, self).__init__()
+
+        mean = [
+            -0.2289, -0.0052, -0.1323, -0.2339, -0.2799,  0.0174,  0.1838,  0.1557,
+            -0.1382,  0.0542,  0.2813,  0.0891,  0.1570, -0.0098,  0.0375, -0.1825,
+            -0.2246, -0.1207, -0.0698,  0.5109,  0.2665, -0.2108, -0.2158,  0.2502,
+            -0.2055, -0.0322,  0.1109,  0.1567, -0.0729,  0.0899, -0.2799, -0.1230,
+            -0.0313, -0.1649,  0.0117,  0.0723, -0.2839, -0.2083, -0.0520,  0.3748,
+            0.0152,  0.1957,  0.1433, -0.2944,  0.3573, -0.0548, -0.1681, -0.0667
+        ]
+        std = [
+            0.4765, 1.0364, 0.4514, 1.1677, 0.5313, 0.4990, 0.4818, 0.5013,
+            0.8158, 1.0344, 0.5894, 1.0901, 0.6885, 0.6165, 0.8454, 0.4978,
+            0.5759, 0.3523, 0.7135, 0.6804, 0.5833, 1.4146, 0.8986, 0.5659,
+            0.7069, 0.5338, 0.4889, 0.4917, 0.4069, 0.4999, 0.6866, 0.4093,
+            0.5709, 0.6065, 0.6415, 0.4944, 0.5726, 1.2042, 0.5458, 1.6887,
+            0.3971, 1.0600, 0.3943, 0.5537, 0.5444, 0.4089, 0.7468, 0.7744
+        ]
+        self.mean = torch.tensor(mean)
+        self.std = torch.tensor(std)
+        self.scale = [self.mean, 1.0 / self.std]
+
+        # init model
+        self.model = VideoVAE38_(z_dim=z_dim, dim=dim).eval().requires_grad_(False)
+        self.upsampling_factor = 16
+        self.z_dim = z_dim
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Diffusers-compatible wrapper (formerly kiwi_vae.py)
+# ─────────────────────────────────��───────────────────────────────────────────
+
+@dataclass
+class LatentDist:
+    mu: torch.Tensor
+
+    def sample(self):
+        return self.mu
+
+
+@dataclass
+class EncoderOutput:
+    latent_dist: LatentDist
+
+
+@dataclass
+class DecoderOutput:
+    sample: torch.Tensor
+
+
+class VAE(VideoVAE_, ModelMixin, ConfigMixin):
+    """
+    Diffusers-compatible VAE wrapper around the original Wan VideoVAE.
+    Loads weights directly from diffusion_pytorch_model.safetensors.
+    """
+
+    @register_to_config
+    def __init__(
+        self,
+        z_dim: int = 48,
+        dim: int = 160,
+        dim_mult: List[int] = [1, 2, 4, 4],
+        num_res_blocks: int = 2,
+        attn_scales: List[float] = [],
+        temperal_downsample: List[bool] = [False, True, True],
+        dropout: float = 0.0,
+        vae_pth: str = "wan_vae.pth",
+        latents_mean: Optional[List[float]] = None,
+        latents_std: Optional[List[float]] = None,
+    ):
+        # Build the actual VAE backbone so diffusers can load weights without mismatch.
+        if z_dim == 48:
+            VideoVAE38_.__init__(
+                self,
+                dim=dim,
+                z_dim=z_dim,
+                dim_mult=dim_mult,
+                num_res_blocks=num_res_blocks,
+                attn_scales=attn_scales,
+                temperal_downsample=temperal_downsample,
+                dropout=dropout,
+            )
+            self._use_38 = True
+            self.upsampling_factor = 16
+        else:
+            VideoVAE_.__init__(
+                self,
+                dim=dim,
+                z_dim=z_dim,
+                dim_mult=dim_mult,
+                num_res_blocks=num_res_blocks,
+                attn_scales=attn_scales,
+                temperal_downsample=temperal_downsample,
+                dropout=dropout,
+            )
+            self._use_38 = False
+            self.upsampling_factor = 8
+
+        # Keep for config compatibility; weights are loaded by diffusers.
+        self._vae_pth = vae_pth
+        self.z_dim = z_dim
+
+        # Build latent normalization scale: [mean, 1/std]
+        if latents_mean is not None and latents_std is not None:
+            mean = torch.tensor(latents_mean)
+            std = torch.tensor(latents_std)
+            self._scale = [mean, 1.0 / std]
+        else:
+            self._scale = [torch.zeros(z_dim), torch.ones(z_dim)]
+
+    def encode(self, x):
+        x = x.to(dtype=next(self.parameters()).dtype)
+        if self._use_38:
+            mu = VideoVAE38_.encode(self, x, self._scale)
+        else:
+            mu = VideoVAE_.encode(self, x, self._scale)
+        return EncoderOutput(latent_dist=LatentDist(mu=mu))
+
+    def decode(self, z):
+        z = z.to(dtype=next(self.parameters()).dtype)
+        if self._use_38:
+            out = VideoVAE38_.decode(self, z, self._scale)
+        else:
+            out = VideoVAE_.decode(self, z, self._scale)
+        return DecoderOutput(sample=out)

wan_video_vae.py

@@ -0,0 +1,1486 @@
+from dataclasses import dataclass
+from typing import List, Optional
+
+from einops import rearrange, repeat
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from tqdm import tqdm
+from diffusers import ModelMixin, ConfigMixin
+from diffusers.configuration_utils import register_to_config
+
+CACHE_T = 2
+
+
+def check_is_instance(model, module_class):
+    if isinstance(model, module_class):
+        return True
+    if hasattr(model, "module") and isinstance(model.module, module_class):
+        return True
+    return False
+
+
+def block_causal_mask(x, block_size):
+    # params
+    b, n, s, _, device = *x.size(), x.device
+    assert s % block_size == 0
+    num_blocks = s // block_size
+
+    # build mask
+    mask = torch.zeros(b, n, s, s, dtype=torch.bool, device=device)
+    for i in range(num_blocks):
+        mask[:, :,
+             i * block_size:(i + 1) * block_size, :(i + 1) * block_size] = 1
+    return mask
+
+
+class CausalConv3d(nn.Conv3d):
+    """
+    Causal 3d convolusion.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._padding = (self.padding[2], self.padding[2], self.padding[1],
+                         self.padding[1], 2 * self.padding[0], 0)
+        self.padding = (0, 0, 0)
+
+    def forward(self, x, cache_x=None):
+        padding = list(self._padding)
+        if cache_x is not None and self._padding[4] > 0:
+            cache_x = cache_x.to(x.device)
+            x = torch.cat([cache_x, x], dim=2)
+            padding[4] -= cache_x.shape[2]
+        x = F.pad(x, padding)
+
+        return super().forward(x)
+
+
+class RMS_norm(nn.Module):
+
+    def __init__(self, dim, channel_first=True, images=True, bias=False):
+        super().__init__()
+        broadcastable_dims = (1, 1, 1) if not images else (1, 1)
+        shape = (dim, *broadcastable_dims) if channel_first else (dim,)
+
+        self.channel_first = channel_first
+        self.scale = dim**0.5
+        self.gamma = nn.Parameter(torch.ones(shape))
+        self.bias = nn.Parameter(torch.zeros(shape)) if bias else 0.
+
+    def forward(self, x):
+        return F.normalize(
+            x, dim=(1 if self.channel_first else
+                    -1)) * self.scale * self.gamma + self.bias
+
+
+class Upsample(nn.Upsample):
+
+    def forward(self, x):
+        """
+        Fix bfloat16 support for nearest neighbor interpolation.
+        """
+        return super().forward(x.float()).type_as(x)
+
+
+class Resample(nn.Module):
+
+    def __init__(self, dim, mode):
+        assert mode in ('none', 'upsample2d', 'upsample3d', 'downsample2d',
+                        'downsample3d')
+        super().__init__()
+        self.dim = dim
+        self.mode = mode
+
+        # layers
+        if mode == 'upsample2d':
+            self.resample = nn.Sequential(
+                Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                nn.Conv2d(dim, dim // 2, 3, padding=1))
+        elif mode == 'upsample3d':
+            self.resample = nn.Sequential(
+                Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                nn.Conv2d(dim, dim // 2, 3, padding=1))
+            self.time_conv = CausalConv3d(dim,
+                                          dim * 2, (3, 1, 1),
+                                          padding=(1, 0, 0))
+
+        elif mode == 'downsample2d':
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                nn.Conv2d(dim, dim, 3, stride=(2, 2)))
+        elif mode == 'downsample3d':
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                nn.Conv2d(dim, dim, 3, stride=(2, 2)))
+            self.time_conv = CausalConv3d(dim,
+                                          dim, (3, 1, 1),
+                                          stride=(2, 1, 1),
+                                          padding=(0, 0, 0))
+
+        else:
+            self.resample = nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        b, c, t, h, w = x.size()
+        if self.mode == 'upsample3d':
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = 'Rep'
+                    feat_idx[0] += 1
+                else:
+
+                    cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                    if cache_x.shape[2] < 2 and feat_cache[
+                            idx] is not None and feat_cache[idx] != 'Rep':
+                        # cache last frame of last two chunk
+                        cache_x = torch.cat([
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device), cache_x
+                        ],
+                                            dim=2)
+                    if cache_x.shape[2] < 2 and feat_cache[
+                            idx] is not None and feat_cache[idx] == 'Rep':
+                        cache_x = torch.cat([
+                            torch.zeros_like(cache_x).to(cache_x.device),
+                            cache_x
+                        ],
+                                            dim=2)
+                    if feat_cache[idx] == 'Rep':
+                        x = self.time_conv(x)
+                    else:
+                        x = self.time_conv(x, feat_cache[idx])
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+
+                    x = x.reshape(b, 2, c, t, h, w)
+                    x = torch.stack((x[:, 0, :, :, :, :], x[:, 1, :, :, :, :]),
+                                    3)
+                    x = x.reshape(b, c, t * 2, h, w)
+        t = x.shape[2]
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+        x = self.resample(x)
+        x = rearrange(x, '(b t) c h w -> b c t h w', t=t)
+
+        if self.mode == 'downsample3d':
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = x.clone()
+                    feat_idx[0] += 1
+                else:
+                    cache_x = x[:, :, -1:, :, :].clone()
+                    x = self.time_conv(
+                        torch.cat([feat_cache[idx][:, :, -1:, :, :], x], 2))
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+        return x
+
+    def init_weight(self, conv):
+        conv_weight = conv.weight
+        nn.init.zeros_(conv_weight)
+        c1, c2, t, h, w = conv_weight.size()
+        one_matrix = torch.eye(c1, c2)
+        init_matrix = one_matrix
+        nn.init.zeros_(conv_weight)
+        conv_weight.data[:, :, 1, 0, 0] = init_matrix
+        conv.weight.data.copy_(conv_weight)
+        nn.init.zeros_(conv.bias.data)
+
+    def init_weight2(self, conv):
+        conv_weight = conv.weight.data
+        nn.init.zeros_(conv_weight)
+        c1, c2, t, h, w = conv_weight.size()
+        init_matrix = torch.eye(c1 // 2, c2)
+        conv_weight[:c1 // 2, :, -1, 0, 0] = init_matrix
+        conv_weight[c1 // 2:, :, -1, 0, 0] = init_matrix
+        conv.weight.data.copy_(conv_weight)
+        nn.init.zeros_(conv.bias.data)
+
+
+
+def patchify(x, patch_size):
+    if patch_size == 1:
+        return x
+    if x.dim() == 4:
+        x = rearrange(x, "b c (h q) (w r) -> b (c r q) h w", q=patch_size, r=patch_size)
+    elif x.dim() == 5:
+        x = rearrange(x,
+                      "b c f (h q) (w r) -> b (c r q) f h w",
+                      q=patch_size,
+                      r=patch_size)
+    else:
+        raise ValueError(f"Invalid input shape: {x.shape}")
+    return x
+
+
+def unpatchify(x, patch_size):
+    if patch_size == 1:
+        return x
+    if x.dim() == 4:
+        x = rearrange(x, "b (c r q) h w -> b c (h q) (w r)", q=patch_size, r=patch_size)
+    elif x.dim() == 5:
+        x = rearrange(x,
+                      "b (c r q) f h w -> b c f (h q) (w r)",
+                      q=patch_size,
+                      r=patch_size)
+    return x
+
+
+class Resample38(Resample):
+
+    def __init__(self, dim, mode):
+        assert mode in (
+            "none",
+            "upsample2d",
+            "upsample3d",
+            "downsample2d",
+            "downsample3d",
+        )
+        super(Resample, self).__init__()
+        self.dim = dim
+        self.mode = mode
+
+        # layers
+        if mode == "upsample2d":
+            self.resample = nn.Sequential(
+                Upsample(scale_factor=(2.0, 2.0), mode="nearest-exact"),
+                nn.Conv2d(dim, dim, 3, padding=1),
+            )
+        elif mode == "upsample3d":
+            self.resample = nn.Sequential(
+                Upsample(scale_factor=(2.0, 2.0), mode="nearest-exact"),
+                nn.Conv2d(dim, dim, 3, padding=1),
+            )
+            self.time_conv = CausalConv3d(dim, dim * 2, (3, 1, 1), padding=(1, 0, 0))
+        elif mode == "downsample2d":
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)), nn.Conv2d(dim, dim, 3, stride=(2, 2))
+            )
+        elif mode == "downsample3d":
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)), nn.Conv2d(dim, dim, 3, stride=(2, 2))
+            )
+            self.time_conv = CausalConv3d(
+                dim, dim, (3, 1, 1), stride=(2, 1, 1), padding=(0, 0, 0)
+            )
+        else:
+            self.resample = nn.Identity()
+
+class ResidualBlock(nn.Module):
+
+    def __init__(self, in_dim, out_dim, dropout=0.0):
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+
+        # layers
+        self.residual = nn.Sequential(
+            RMS_norm(in_dim, images=False), nn.SiLU(),
+            CausalConv3d(in_dim, out_dim, 3, padding=1),
+            RMS_norm(out_dim, images=False), nn.SiLU(), nn.Dropout(dropout),
+            CausalConv3d(out_dim, out_dim, 3, padding=1))
+        self.shortcut = CausalConv3d(in_dim, out_dim, 1) \
+            if in_dim != out_dim else nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        h = self.shortcut(x)
+        for layer in self.residual:
+            if check_is_instance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x + h
+
+
+class AttentionBlock(nn.Module):
+    """
+    Causal self-attention with a single head.
+    """
+
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+
+        # layers
+        self.norm = RMS_norm(dim)
+        self.to_qkv = nn.Conv2d(dim, dim * 3, 1)
+        self.proj = nn.Conv2d(dim, dim, 1)
+
+        # zero out the last layer params
+        nn.init.zeros_(self.proj.weight)
+
+    def forward(self, x):
+        identity = x
+        b, c, t, h, w = x.size()
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+        x = self.norm(x)
+        # compute query, key, value
+        q, k, v = self.to_qkv(x).reshape(b * t, 1, c * 3, -1).permute(
+            0, 1, 3, 2).contiguous().chunk(3, dim=-1)
+
+        # apply attention
+        x = F.scaled_dot_product_attention(
+            q,
+            k,
+            v,
+            #attn_mask=block_causal_mask(q, block_size=h * w)
+        )
+        x = x.squeeze(1).permute(0, 2, 1).reshape(b * t, c, h, w)
+
+        # output
+        x = self.proj(x)
+        x = rearrange(x, '(b t) c h w-> b c t h w', t=t)
+        return x + identity
+
+
+class AvgDown3D(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        factor_t,
+        factor_s=1,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.factor_t = factor_t
+        self.factor_s = factor_s
+        self.factor = self.factor_t * self.factor_s * self.factor_s
+
+        assert in_channels * self.factor % out_channels == 0
+        self.group_size = in_channels * self.factor // out_channels
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        pad_t = (self.factor_t - x.shape[2] % self.factor_t) % self.factor_t
+        pad = (0, 0, 0, 0, pad_t, 0)
+        x = F.pad(x, pad)
+        B, C, T, H, W = x.shape
+        x = x.view(
+            B,
+            C,
+            T // self.factor_t,
+            self.factor_t,
+            H // self.factor_s,
+            self.factor_s,
+            W // self.factor_s,
+            self.factor_s,
+        )
+        x = x.permute(0, 1, 3, 5, 7, 2, 4, 6).contiguous()
+        x = x.view(
+            B,
+            C * self.factor,
+            T // self.factor_t,
+            H // self.factor_s,
+            W // self.factor_s,
+        )
+        x = x.view(
+            B,
+            self.out_channels,
+            self.group_size,
+            T // self.factor_t,
+            H // self.factor_s,
+            W // self.factor_s,
+        )
+        x = x.mean(dim=2)
+        return x
+
+
+class DupUp3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        factor_t,
+        factor_s=1,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        self.factor_t = factor_t
+        self.factor_s = factor_s
+        self.factor = self.factor_t * self.factor_s * self.factor_s
+
+        assert out_channels * self.factor % in_channels == 0
+        self.repeats = out_channels * self.factor // in_channels
+
+    def forward(self, x: torch.Tensor, first_chunk=False) -> torch.Tensor:
+        x = x.repeat_interleave(self.repeats, dim=1)
+        x = x.view(
+            x.size(0),
+            self.out_channels,
+            self.factor_t,
+            self.factor_s,
+            self.factor_s,
+            x.size(2),
+            x.size(3),
+            x.size(4),
+        )
+        x = x.permute(0, 1, 5, 2, 6, 3, 7, 4).contiguous()
+        x = x.view(
+            x.size(0),
+            self.out_channels,
+            x.size(2) * self.factor_t,
+            x.size(4) * self.factor_s,
+            x.size(6) * self.factor_s,
+        )
+        if first_chunk:
+            x = x[:, :, self.factor_t - 1 :, :, :]
+        return x
+
+
+class Down_ResidualBlock(nn.Module):
+    def __init__(
+        self, in_dim, out_dim, dropout, mult, temperal_downsample=False, down_flag=False
+    ):
+        super().__init__()
+
+        # Shortcut path with downsample
+        self.avg_shortcut = AvgDown3D(
+            in_dim,
+            out_dim,
+            factor_t=2 if temperal_downsample else 1,
+            factor_s=2 if down_flag else 1,
+        )
+
+        # Main path with residual blocks and downsample
+        downsamples = []
+        for _ in range(mult):
+            downsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+            in_dim = out_dim
+
+        # Add the final downsample block
+        if down_flag:
+            mode = "downsample3d" if temperal_downsample else "downsample2d"
+            downsamples.append(Resample38(out_dim, mode=mode))
+
+        self.downsamples = nn.Sequential(*downsamples)
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        x_copy = x.clone()
+        for module in self.downsamples:
+            x = module(x, feat_cache, feat_idx)
+
+        return x + self.avg_shortcut(x_copy)
+
+
+class Up_ResidualBlock(nn.Module):
+    def __init__(
+        self, in_dim, out_dim, dropout, mult, temperal_upsample=False, up_flag=False
+    ):
+        super().__init__()
+        # Shortcut path with upsample
+        if up_flag:
+            self.avg_shortcut = DupUp3D(
+                in_dim,
+                out_dim,
+                factor_t=2 if temperal_upsample else 1,
+                factor_s=2 if up_flag else 1,
+            )
+        else:
+            self.avg_shortcut = None
+
+        # Main path with residual blocks and upsample
+        upsamples = []
+        for _ in range(mult):
+            upsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+            in_dim = out_dim
+
+        # Add the final upsample block
+        if up_flag:
+            mode = "upsample3d" if temperal_upsample else "upsample2d"
+            upsamples.append(Resample38(out_dim, mode=mode))
+
+        self.upsamples = nn.Sequential(*upsamples)
+
+    def forward(self, x, feat_cache=None, feat_idx=[0], first_chunk=False):
+        x_main = x.clone()
+        for module in self.upsamples:
+            x_main = module(x_main, feat_cache, feat_idx)
+        if self.avg_shortcut is not None:
+            x_shortcut = self.avg_shortcut(x, first_chunk)
+            return x_main + x_shortcut
+        else:
+            return x_main
+
+
+class Encoder3d(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[True, True, False],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+
+        # dimensions
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        # init block
+        self.conv1 = CausalConv3d(3, dims[0], 3, padding=1)
+
+        # downsample blocks
+        downsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            for _ in range(num_res_blocks):
+                downsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    downsamples.append(AttentionBlock(out_dim))
+                in_dim = out_dim
+
+            # downsample block
+            if i != len(dim_mult) - 1:
+                mode = 'downsample3d' if temperal_downsample[
+                    i] else 'downsample2d'
+                downsamples.append(Resample(out_dim, mode=mode))
+                scale /= 2.0
+        self.downsamples = nn.Sequential(*downsamples)
+
+        # middle blocks
+        self.middle = nn.Sequential(ResidualBlock(out_dim, out_dim, dropout),
+                                    AttentionBlock(out_dim),
+                                    ResidualBlock(out_dim, out_dim, dropout))
+
+        # output blocks
+        self.head = nn.Sequential(RMS_norm(out_dim, images=False), nn.SiLU(),
+                                  CausalConv3d(out_dim, z_dim, 3, padding=1))
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([
+                    feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                        cache_x.device), cache_x
+                ],
+                                    dim=2)
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## downsamples
+        for layer in self.downsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## middle
+        for layer in self.middle:
+            if check_is_instance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if check_is_instance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+class Encoder3d_38(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[False, True, True],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+
+        # dimensions
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        # init block
+        self.conv1 = CausalConv3d(12, dims[0], 3, padding=1)
+
+        # downsample blocks
+        downsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            t_down_flag = (
+                temperal_downsample[i] if i < len(temperal_downsample) else False
+            )
+            downsamples.append(
+                Down_ResidualBlock(
+                    in_dim=in_dim,
+                    out_dim=out_dim,
+                    dropout=dropout,
+                    mult=num_res_blocks,
+                    temperal_downsample=t_down_flag,
+                    down_flag=i != len(dim_mult) - 1,
+                )
+            )
+            scale /= 2.0
+        self.downsamples = nn.Sequential(*downsamples)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(out_dim, out_dim, dropout),
+            AttentionBlock(out_dim),
+            ResidualBlock(out_dim, out_dim, dropout),
+        )
+
+        # # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False),
+            nn.SiLU(),
+            CausalConv3d(out_dim, z_dim, 3, padding=1),
+        )
+
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device),
+                        cache_x,
+                    ],
+                    dim=2,
+                )
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## downsamples
+        for layer in self.downsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## middle
+        for layer in self.middle:
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    cache_x = torch.cat(
+                        [
+                            feat_cache[idx][:, :, -1, :, :]
+                            .unsqueeze(2)
+                            .to(cache_x.device),
+                            cache_x,
+                        ],
+                        dim=2,
+                    )
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+
+        return x
+
+
+class Decoder3d(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_upsample=[False, True, True],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_upsample = temperal_upsample
+
+        # dimensions
+        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
+        scale = 1.0 / 2**(len(dim_mult) - 2)
+
+        # init block
+        self.conv1 = CausalConv3d(z_dim, dims[0], 3, padding=1)
+
+        # middle blocks
+        self.middle = nn.Sequential(ResidualBlock(dims[0], dims[0], dropout),
+                                    AttentionBlock(dims[0]),
+                                    ResidualBlock(dims[0], dims[0], dropout))
+
+        # upsample blocks
+        upsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            if i == 1 or i == 2 or i == 3:
+                in_dim = in_dim // 2
+            for _ in range(num_res_blocks + 1):
+                upsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    upsamples.append(AttentionBlock(out_dim))
+                in_dim = out_dim
+
+            # upsample block
+            if i != len(dim_mult) - 1:
+                mode = 'upsample3d' if temperal_upsample[i] else 'upsample2d'
+                upsamples.append(Resample(out_dim, mode=mode))
+                scale *= 2.0
+        self.upsamples = nn.Sequential(*upsamples)
+
+        # output blocks
+        self.head = nn.Sequential(RMS_norm(out_dim, images=False), nn.SiLU(),
+                                  CausalConv3d(out_dim, 3, 3, padding=1))
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        ## conv1
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([
+                    feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                        cache_x.device), cache_x
+                ],
+                                    dim=2)
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## middle
+        for layer in self.middle:
+            if check_is_instance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## upsamples
+        for layer in self.upsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if check_is_instance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+
+class Decoder3d_38(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_upsample=[False, True, True],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_upsample = temperal_upsample
+
+        # dimensions
+        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
+        scale = 1.0 / 2 ** (len(dim_mult) - 2)
+        # init block
+        self.conv1 = CausalConv3d(z_dim, dims[0], 3, padding=1)
+
+        # middle blocks
+        self.middle = nn.Sequential(ResidualBlock(dims[0], dims[0], dropout),
+                                    AttentionBlock(dims[0]),
+                                    ResidualBlock(dims[0], dims[0], dropout))
+
+        # upsample blocks
+        upsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            t_up_flag = temperal_upsample[i] if i < len(temperal_upsample) else False
+            upsamples.append(
+                Up_ResidualBlock(in_dim=in_dim,
+                                 out_dim=out_dim,
+                                 dropout=dropout,
+                                 mult=num_res_blocks + 1,
+                                 temperal_upsample=t_up_flag,
+                                 up_flag=i != len(dim_mult) - 1))
+        self.upsamples = nn.Sequential(*upsamples)
+
+        # output blocks
+        self.head = nn.Sequential(RMS_norm(out_dim, images=False), nn.SiLU(),
+                                  CausalConv3d(out_dim, 12, 3, padding=1))
+
+
+    def forward(self, x, feat_cache=None, feat_idx=[0], first_chunk=False):
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device),
+                        cache_x,
+                    ],
+                    dim=2,
+                )
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        for layer in self.middle:
+            if check_is_instance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## upsamples
+        for layer in self.upsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx, first_chunk)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if check_is_instance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    cache_x = torch.cat(
+                        [
+                            feat_cache[idx][:, :, -1, :, :]
+                            .unsqueeze(2)
+                            .to(cache_x.device),
+                            cache_x,
+                        ],
+                        dim=2,
+                    )
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+def count_conv3d(model):
+    count = 0
+    for m in model.modules():
+        if isinstance(m, CausalConv3d):
+            count += 1
+    return count
+
+
+class VideoVAE_(nn.Module):
+
+    def __init__(self,
+                 dim=96,
+                 z_dim=16,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[False, True, True],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+        self.temperal_upsample = temperal_downsample[::-1]
+
+        # modules
+        self.encoder = Encoder3d(dim, z_dim * 2, dim_mult, num_res_blocks,
+                                 attn_scales, self.temperal_downsample, dropout)
+        self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
+        self.conv2 = CausalConv3d(z_dim, z_dim, 1)
+        self.decoder = Decoder3d(dim, z_dim, dim_mult, num_res_blocks,
+                                 attn_scales, self.temperal_upsample, dropout)
+
+    def forward(self, x):
+        mu, log_var = self.encode(x)
+        z = self.reparameterize(mu, log_var)
+        x_recon = self.decode(z)
+        return x_recon, mu, log_var
+
+    def encode(self, x, scale):
+        self.clear_cache()
+        ## cache
+        t = x.shape[2]
+        iter_ = 1 + (t - 1) // 4
+
+        for i in range(iter_):
+            self._enc_conv_idx = [0]
+            if i == 0:
+                out = self.encoder(x[:, :, :1, :, :],
+                                   feat_cache=self._enc_feat_map,
+                                   feat_idx=self._enc_conv_idx)
+            else:
+                out_ = self.encoder(x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
+                                    feat_cache=self._enc_feat_map,
+                                    feat_idx=self._enc_conv_idx)
+                out = torch.cat([out, out_], 2)
+        mu, log_var = self.conv1(out).chunk(2, dim=1)
+        if isinstance(scale[0], torch.Tensor):
+            scale = [s.to(dtype=mu.dtype, device=mu.device) for s in scale]
+            mu = (mu - scale[0].view(1, self.z_dim, 1, 1, 1)) * scale[1].view(
+                1, self.z_dim, 1, 1, 1)
+        else:
+            scale = scale.to(dtype=mu.dtype, device=mu.device)
+            mu = (mu - scale[0]) * scale[1]
+        return mu
+
+    def decode(self, z, scale):
+        self.clear_cache()
+        # z: [b,c,t,h,w]
+        if isinstance(scale[0], torch.Tensor):
+            scale = [s.to(dtype=z.dtype, device=z.device) for s in scale]
+            z = z / scale[1].view(1, self.z_dim, 1, 1, 1) + scale[0].view(
+                1, self.z_dim, 1, 1, 1)
+        else:
+            scale = scale.to(dtype=z.dtype, device=z.device)
+            z = z / scale[1] + scale[0]
+        iter_ = z.shape[2]
+        x = self.conv2(z)
+        for i in range(iter_):
+            self._conv_idx = [0]
+            if i == 0:
+                out = self.decoder(x[:, :, i:i + 1, :, :],
+                                   feat_cache=self._feat_map,
+                                   feat_idx=self._conv_idx)
+            else:
+                out_ = self.decoder(x[:, :, i:i + 1, :, :],
+                                    feat_cache=self._feat_map,
+                                    feat_idx=self._conv_idx)
+                out = torch.cat([out, out_], 2) # may add tensor offload
+        return out
+
+    def reparameterize(self, mu, log_var):
+        std = torch.exp(0.5 * log_var)
+        eps = torch.randn_like(std)
+        return eps * std + mu
+
+    def sample(self, imgs, deterministic=False):
+        mu, log_var = self.encode(imgs)
+        if deterministic:
+            return mu
+        std = torch.exp(0.5 * log_var.clamp(-30.0, 20.0))
+        return mu + std * torch.randn_like(std)
+
+    def clear_cache(self):
+        self._conv_num = count_conv3d(self.decoder)
+        self._conv_idx = [0]
+        self._feat_map = [None] * self._conv_num
+        # cache encode
+        self._enc_conv_num = count_conv3d(self.encoder)
+        self._enc_conv_idx = [0]
+        self._enc_feat_map = [None] * self._enc_conv_num
+
+
+class WanVideoVAE(nn.Module):
+
+    def __init__(self, z_dim=16):
+        super().__init__()
+
+        mean = [
+            -0.7571, -0.7089, -0.9113, 0.1075, -0.1745, 0.9653, -0.1517, 1.5508,
+            0.4134, -0.0715, 0.5517, -0.3632, -0.1922, -0.9497, 0.2503, -0.2921
+        ]
+        std = [
+            2.8184, 1.4541, 2.3275, 2.6558, 1.2196, 1.7708, 2.6052, 2.0743,
+            3.2687, 2.1526, 2.8652, 1.5579, 1.6382, 1.1253, 2.8251, 1.9160
+        ]
+        self.mean = torch.tensor(mean)
+        self.std = torch.tensor(std)
+        self.scale = [self.mean, 1.0 / self.std]
+
+        # init model
+        self.model = VideoVAE_(z_dim=z_dim).eval().requires_grad_(False)
+        self.upsampling_factor = 8
+        self.z_dim = z_dim
+
+
+    def build_1d_mask(self, length, left_bound, right_bound, border_width):
+        x = torch.ones((length,))
+        if not left_bound:
+            x[:border_width] = (torch.arange(border_width) + 1) / border_width
+        if not right_bound:
+            x[-border_width:] = torch.flip((torch.arange(border_width) + 1) / border_width, dims=(0,))
+        return x
+
+
+    def build_mask(self, data, is_bound, border_width):
+        _, _, _, H, W = data.shape
+        h = self.build_1d_mask(H, is_bound[0], is_bound[1], border_width[0])
+        w = self.build_1d_mask(W, is_bound[2], is_bound[3], border_width[1])
+
+        h = repeat(h, "H -> H W", H=H, W=W)
+        w = repeat(w, "W -> H W", H=H, W=W)
+
+        mask = torch.stack([h, w]).min(dim=0).values
+        mask = rearrange(mask, "H W -> 1 1 1 H W")
+        return mask
+
+
+    def tiled_decode(self, hidden_states, device, tile_size, tile_stride):
+        _, _, T, H, W = hidden_states.shape
+        size_h, size_w = tile_size
+        stride_h, stride_w = tile_stride
+
+        # Split tasks
+        tasks = []
+        for h in range(0, H, stride_h):
+            if (h-stride_h >= 0 and h-stride_h+size_h >= H): continue
+            for w in range(0, W, stride_w):
+                if (w-stride_w >= 0 and w-stride_w+size_w >= W): continue
+                h_, w_ = h + size_h, w + size_w
+                tasks.append((h, h_, w, w_))
+
+        data_device = "cpu"
+        computation_device = device
+
+        out_T = T * 4 - 3
+        weight = torch.zeros((1, 1, out_T, H * self.upsampling_factor, W * self.upsampling_factor), dtype=hidden_states.dtype, device=data_device)
+        values = torch.zeros((1, 3, out_T, H * self.upsampling_factor, W * self.upsampling_factor), dtype=hidden_states.dtype, device=data_device)
+
+        for h, h_, w, w_ in tqdm(tasks, desc="VAE decoding"):
+            hidden_states_batch = hidden_states[:, :, :, h:h_, w:w_].to(computation_device)
+            hidden_states_batch = self.model.decode(hidden_states_batch, self.scale).to(data_device)
+
+            mask = self.build_mask(
+                hidden_states_batch,
+                is_bound=(h==0, h_>=H, w==0, w_>=W),
+                border_width=((size_h - stride_h) * self.upsampling_factor, (size_w - stride_w) * self.upsampling_factor)
+            ).to(dtype=hidden_states.dtype, device=data_device)
+
+            target_h = h * self.upsampling_factor
+            target_w = w * self.upsampling_factor
+            values[
+                :,
+                :,
+                :,
+                target_h:target_h + hidden_states_batch.shape[3],
+                target_w:target_w + hidden_states_batch.shape[4],
+            ] += hidden_states_batch * mask
+            weight[
+                :,
+                :,
+                :,
+                target_h: target_h + hidden_states_batch.shape[3],
+                target_w: target_w + hidden_states_batch.shape[4],
+            ] += mask
+        values = values / weight
+        values = values.clamp_(-1, 1)
+        return values
+
+
+    def tiled_encode(self, video, device, tile_size, tile_stride):
+        _, _, T, H, W = video.shape
+        size_h, size_w = tile_size
+        stride_h, stride_w = tile_stride
+
+        # Split tasks
+        tasks = []
+        for h in range(0, H, stride_h):
+            if (h-stride_h >= 0 and h-stride_h+size_h >= H): continue
+            for w in range(0, W, stride_w):
+                if (w-stride_w >= 0 and w-stride_w+size_w >= W): continue
+                h_, w_ = h + size_h, w + size_w
+                tasks.append((h, h_, w, w_))
+
+        data_device = "cpu"
+        computation_device = device
+
+        out_T = (T + 3) // 4
+        weight = torch.zeros((1, 1, out_T, H // self.upsampling_factor, W // self.upsampling_factor), dtype=video.dtype, device=data_device)
+        values = torch.zeros((1, self.z_dim, out_T, H // self.upsampling_factor, W // self.upsampling_factor), dtype=video.dtype, device=data_device)
+
+        for h, h_, w, w_ in tqdm(tasks, desc="VAE encoding"):
+            hidden_states_batch = video[:, :, :, h:h_, w:w_].to(computation_device)
+            hidden_states_batch = self.model.encode(hidden_states_batch, self.scale).to(data_device)
+
+            mask = self.build_mask(
+                hidden_states_batch,
+                is_bound=(h==0, h_>=H, w==0, w_>=W),
+                border_width=((size_h - stride_h) // self.upsampling_factor, (size_w - stride_w) // self.upsampling_factor)
+            ).to(dtype=video.dtype, device=data_device)
+
+            target_h = h // self.upsampling_factor
+            target_w = w // self.upsampling_factor
+            values[
+                :,
+                :,
+                :,
+                target_h:target_h + hidden_states_batch.shape[3],
+                target_w:target_w + hidden_states_batch.shape[4],
+            ] += hidden_states_batch * mask
+            weight[
+                :,
+                :,
+                :,
+                target_h: target_h + hidden_states_batch.shape[3],
+                target_w: target_w + hidden_states_batch.shape[4],
+            ] += mask
+        values = values / weight
+        return values
+
+
+    def single_encode(self, video, device):
+        video = video.to(device)
+        x = self.model.encode(video, self.scale)
+        return x
+
+
+    def single_decode(self, hidden_state, device):
+        hidden_state = hidden_state.to(device)
+        video = self.model.decode(hidden_state, self.scale)
+        return video.clamp_(-1, 1)
+
+
+    def encode(self, videos, device, tiled=False, tile_size=(34, 34), tile_stride=(18, 16)):
+        videos = [video.to("cpu") for video in videos]
+        hidden_states = []
+        for video in videos:
+            video = video.unsqueeze(0)
+            if tiled:
+                tile_size = (tile_size[0] * self.upsampling_factor, tile_size[1] * self.upsampling_factor)
+                tile_stride = (tile_stride[0] * self.upsampling_factor, tile_stride[1] * self.upsampling_factor)
+                hidden_state = self.tiled_encode(video, device, tile_size, tile_stride)
+            else:
+                hidden_state = self.single_encode(video, device)
+            hidden_state = hidden_state.squeeze(0)
+            hidden_states.append(hidden_state)
+        hidden_states = torch.stack(hidden_states)
+        return hidden_states
+
+
+    def decode(self, hidden_states, device, tiled=False, tile_size=(34, 34), tile_stride=(18, 16)):
+        hidden_states = [hidden_state.to("cpu") for hidden_state in hidden_states]
+        videos = []
+        for hidden_state in hidden_states:
+            hidden_state = hidden_state.unsqueeze(0)
+            if tiled:
+                video = self.tiled_decode(hidden_state, device, tile_size, tile_stride)
+            else:
+                video = self.single_decode(hidden_state, device)
+            video = video.squeeze(0)
+            videos.append(video)
+        videos = torch.stack(videos)
+        return videos
+
+
+    @staticmethod
+    def state_dict_converter():
+        return WanVideoVAEStateDictConverter()
+
+
+class WanVideoVAEStateDictConverter:
+
+    def __init__(self):
+        pass
+
+    def from_civitai(self, state_dict):
+        state_dict_ = {}
+        if 'model_state' in state_dict:
+            state_dict = state_dict['model_state']
+        for name in state_dict:
+            state_dict_['model.' + name] = state_dict[name]
+        return state_dict_
+
+
+class VideoVAE38_(VideoVAE_):
+
+    def __init__(self,
+                 dim=160,
+                 z_dim=48,
+                 dec_dim=256,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[False, True, True],
+                 dropout=0.0):
+        super(VideoVAE_, self).__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+        self.temperal_upsample = temperal_downsample[::-1]
+
+        # modules
+        self.encoder = Encoder3d_38(dim, z_dim * 2, dim_mult, num_res_blocks,
+                                    attn_scales, self.temperal_downsample, dropout)
+        self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
+        self.conv2 = CausalConv3d(z_dim, z_dim, 1)
+        self.decoder = Decoder3d_38(dec_dim, z_dim, dim_mult, num_res_blocks,
+                                    attn_scales, self.temperal_upsample, dropout)
+
+
+    def encode(self, x, scale):
+        self.clear_cache()
+        x = patchify(x, patch_size=2)
+        t = x.shape[2]
+        iter_ = 1 + (t - 1) // 4
+        for i in range(iter_):
+            self._enc_conv_idx = [0]
+            if i == 0:
+                out = self.encoder(x[:, :, :1, :, :],
+                                   feat_cache=self._enc_feat_map,
+                                   feat_idx=self._enc_conv_idx)
+            else:
+                out_ = self.encoder(x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
+                                    feat_cache=self._enc_feat_map,
+                                    feat_idx=self._enc_conv_idx)
+                out = torch.cat([out, out_], 2)
+        mu, log_var = self.conv1(out).chunk(2, dim=1)
+        if isinstance(scale[0], torch.Tensor):
+            scale = [s.to(dtype=mu.dtype, device=mu.device) for s in scale]
+            mu = (mu - scale[0].view(1, self.z_dim, 1, 1, 1)) * scale[1].view(
+                1, self.z_dim, 1, 1, 1)
+        else:
+            scale = scale.to(dtype=mu.dtype, device=mu.device)
+            mu = (mu - scale[0]) * scale[1]
+        self.clear_cache()
+        return mu
+
+
+    def decode(self, z, scale):
+        self.clear_cache()
+        if isinstance(scale[0], torch.Tensor):
+            scale = [s.to(dtype=z.dtype, device=z.device) for s in scale]
+            z = z / scale[1].view(1, self.z_dim, 1, 1, 1) + scale[0].view(
+                1, self.z_dim, 1, 1, 1)
+        else:
+            scale = scale.to(dtype=z.dtype, device=z.device)
+            z = z / scale[1] + scale[0]
+        iter_ = z.shape[2]
+        x = self.conv2(z)
+        for i in range(iter_):
+            self._conv_idx = [0]
+            if i == 0:
+                out = self.decoder(x[:, :, i:i + 1, :, :],
+                                   feat_cache=self._feat_map,
+                                   feat_idx=self._conv_idx,
+                                   first_chunk=True)
+            else:
+                out_ = self.decoder(x[:, :, i:i + 1, :, :],
+                                    feat_cache=self._feat_map,
+                                    feat_idx=self._conv_idx)
+                out = torch.cat([out, out_], 2)
+        out = unpatchify(out, patch_size=2)
+        self.clear_cache()
+        return out
+
+
+class WanVideoVAE38(WanVideoVAE):
+
+    def __init__(self, z_dim=48, dim=160):
+        super(WanVideoVAE, self).__init__()
+
+        mean = [
+            -0.2289, -0.0052, -0.1323, -0.2339, -0.2799,  0.0174,  0.1838,  0.1557,
+            -0.1382,  0.0542,  0.2813,  0.0891,  0.1570, -0.0098,  0.0375, -0.1825,
+            -0.2246, -0.1207, -0.0698,  0.5109,  0.2665, -0.2108, -0.2158,  0.2502,
+            -0.2055, -0.0322,  0.1109,  0.1567, -0.0729,  0.0899, -0.2799, -0.1230,
+            -0.0313, -0.1649,  0.0117,  0.0723, -0.2839, -0.2083, -0.0520,  0.3748,
+            0.0152,  0.1957,  0.1433, -0.2944,  0.3573, -0.0548, -0.1681, -0.0667
+        ]
+        std = [
+            0.4765, 1.0364, 0.4514, 1.1677, 0.5313, 0.4990, 0.4818, 0.5013,
+            0.8158, 1.0344, 0.5894, 1.0901, 0.6885, 0.6165, 0.8454, 0.4978,
+            0.5759, 0.3523, 0.7135, 0.6804, 0.5833, 1.4146, 0.8986, 0.5659,
+            0.7069, 0.5338, 0.4889, 0.4917, 0.4069, 0.4999, 0.6866, 0.4093,
+            0.5709, 0.6065, 0.6415, 0.4944, 0.5726, 1.2042, 0.5458, 1.6887,
+            0.3971, 1.0600, 0.3943, 0.5537, 0.5444, 0.4089, 0.7468, 0.7744
+        ]
+        self.mean = torch.tensor(mean)
+        self.std = torch.tensor(std)
+        self.scale = [self.mean, 1.0 / self.std]
+
+        # init model
+        self.model = VideoVAE38_(z_dim=z_dim, dim=dim).eval().requires_grad_(False)
+        self.upsampling_factor = 16
+        self.z_dim = z_dim
+
+
+# ─────────────────────────────────────────────────────────────────────────────
+# Diffusers-compatible wrapper (formerly kiwi_vae.py)
+# ─────────────────────────────────��───────────────────────────────────────────
+
+@dataclass
+class LatentDist:
+    mu: torch.Tensor
+
+    def sample(self):
+        return self.mu
+
+
+@dataclass
+class EncoderOutput:
+    latent_dist: LatentDist
+
+
+@dataclass
+class DecoderOutput:
+    sample: torch.Tensor
+
+
+class VAE(VideoVAE_, ModelMixin, ConfigMixin):
+    """
+    Diffusers-compatible VAE wrapper around the original Wan VideoVAE.
+    Loads weights directly from diffusion_pytorch_model.safetensors.
+    """
+
+    @register_to_config
+    def __init__(
+        self,
+        z_dim: int = 48,
+        dim: int = 160,
+        dim_mult: List[int] = [1, 2, 4, 4],
+        num_res_blocks: int = 2,
+        attn_scales: List[float] = [],
+        temperal_downsample: List[bool] = [False, True, True],
+        dropout: float = 0.0,
+        vae_pth: str = "wan_vae.pth",
+        latents_mean: Optional[List[float]] = None,
+        latents_std: Optional[List[float]] = None,
+    ):
+        # Build the actual VAE backbone so diffusers can load weights without mismatch.
+        if z_dim == 48:
+            VideoVAE38_.__init__(
+                self,
+                dim=dim,
+                z_dim=z_dim,
+                dim_mult=dim_mult,
+                num_res_blocks=num_res_blocks,
+                attn_scales=attn_scales,
+                temperal_downsample=temperal_downsample,
+                dropout=dropout,
+            )
+            self._use_38 = True
+            self.upsampling_factor = 16
+        else:
+            VideoVAE_.__init__(
+                self,
+                dim=dim,
+                z_dim=z_dim,
+                dim_mult=dim_mult,
+                num_res_blocks=num_res_blocks,
+                attn_scales=attn_scales,
+                temperal_downsample=temperal_downsample,
+                dropout=dropout,
+            )
+            self._use_38 = False
+            self.upsampling_factor = 8
+
+        # Keep for config compatibility; weights are loaded by diffusers.
+        self._vae_pth = vae_pth
+        self.z_dim = z_dim
+
+        # Build latent normalization scale: [mean, 1/std]
+        if latents_mean is not None and latents_std is not None:
+            mean = torch.tensor(latents_mean)
+            std = torch.tensor(latents_std)
+            self._scale = [mean, 1.0 / std]
+        else:
+            self._scale = [torch.zeros(z_dim), torch.ones(z_dim)]
+
+    def encode(self, x):
+        x = x.to(dtype=next(self.parameters()).dtype)
+        if self._use_38:
+            mu = VideoVAE38_.encode(self, x, self._scale)
+        else:
+            mu = VideoVAE_.encode(self, x, self._scale)
+        return EncoderOutput(latent_dist=LatentDist(mu=mu))
+
+    def decode(self, z):
+        z = z.to(dtype=next(self.parameters()).dtype)
+        if self._use_38:
+            out = VideoVAE38_.decode(self, z, self._scale)
+        else:
+            out = VideoVAE_.decode(self, z, self._scale)
+        return DecoderOutput(sample=out)