Create app.py

app.py

@@ -0,0 +1,635 @@
+import os
+import subprocess
+import sys
+import json
+import struct
+
+# Disable torch.compile / dynamo before any torch import
+os.environ["TORCH_COMPILE_DISABLE"] = "1"
+os.environ["TORCHDYNAMO_DISABLE"] = "1"
+
+
+# Clone LTX-2 repo and install packages
+LTX_REPO_URL = "https://github.com/Lightricks/LTX-2.git"
+LTX_REPO_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "LTX-2")
+
+LTX_COMMIT = "ae855f8538843825f9015a419cf4ba5edaf5eec2"  # known working commit with decode_video
+
+if not os.path.exists(LTX_REPO_DIR):
+    print(f"Cloning {LTX_REPO_URL}...")
+    subprocess.run(["git", "clone", LTX_REPO_URL, LTX_REPO_DIR], check=True)
+    subprocess.run(["git", "checkout", LTX_COMMIT], cwd=LTX_REPO_DIR, check=True)
+
+print("Installing ltx-core and ltx-pipelines from cloned repo...")
+subprocess.run(
+    [sys.executable, "-m", "pip", "install", "--force-reinstall", "--no-deps", "-e",
+     os.path.join(LTX_REPO_DIR, "packages", "ltx-core"),
+     "-e", os.path.join(LTX_REPO_DIR, "packages", "ltx-pipelines")],
+    check=True,
+)
+
+sys.path.insert(0, os.path.join(LTX_REPO_DIR, "packages", "ltx-pipelines", "src"))
+sys.path.insert(0, os.path.join(LTX_REPO_DIR, "packages", "ltx-core", "src"))
+
+import logging
+import random
+import tempfile
+from pathlib import Path
+import gc
+import hashlib
+import shutil
+
+import spaces
+import torch
+
+torch._dynamo.config.suppress_errors = True
+torch._dynamo.config.disable = True
+
+# --- CRITICAL FIX: ZERO-GPU LOAD PATCH START ---
+from ltx_core.loader.primitives import StateDict
+from ltx_core.loader.sft_loader import SafetensorsStateDictLoader
+
+_SAFETENSORS_DTYPE_MAP = {
+    "F64": torch.float64,
+    "F32": torch.float32,
+    "F16": torch.float16,
+    "BF16": torch.bfloat16,
+    "F8_E5M2": torch.float8_e5m2,
+    "F8_E4M3": torch.float8_e4m3fn,
+    "I64": torch.int64,
+    "I32": torch.int32,
+    "I16": torch.int16,
+    "I8": torch.int8,
+    "U8": torch.uint8,
+    "BOOL": torch.bool,
+}
+
+def _patched_load(self, path, sd_ops, device=None):
+    """
+    Forces tensors to load onto CPU during the startup phase to prevent 
+    'No CUDA GPUs are available' errors in ZeroGPU.
+    """
+    sd = {}
+    size = 0
+    dtype = set()
+    # FORCE CPU during preloading
+    device = torch.device("cpu") 
+    model_paths = path if isinstance(path, list) else [path]
+    for shard_path in model_paths:
+        with open(shard_path, "rb") as f:
+            header_len = struct.unpack("<Q", f.read(8))[0]
+            header = json.loads(f.read(header_len).decode("utf-8"))
+            data_base = 8 + header_len
+            for name, meta in header.items():
+                if name == "__metadata__":
+                    continue
+                expected_name = name if sd_ops is None else sd_ops.apply_to_key(name)
+                if expected_name is None:
+                    continue
+                start, end = meta["data_offsets"]
+                f.seek(data_base + start)
+                buf = f.read(end - start)
+                t = torch.frombuffer(
+                    bytearray(buf), dtype=_SAFETENSORS_DTYPE_MAP[meta["dtype"]]
+                ).reshape(meta["shape"])
+                t = t.to(device=device, non_blocking=True, copy=False)
+                kvs = (
+                    ((expected_name, t),)
+                    if sd_ops is None
+                    else sd_ops.apply_to_key_value(expected_name, t)
+                )
+                for key, v in kvs:
+                    size += v.nbytes
+                    dtype.add(v.dtype)
+                    sd[key] = v
+    return StateDict(sd=sd, device=device, size=size, dtype=dtype)
+
+SafetensorsStateDictLoader.load = _patched_load
+print("[FIX] SafetensorsStateDictLoader.load patched for ZeroGPU")
+# --- CRITICAL FIX END ---
+
+_original_tensor_to = torch.Tensor.to
+
+
+def _is_cuda_target(x):
+    return (
+        x == "cuda"
+        or (isinstance(x, torch.device) and x.type == "cuda")
+        or (isinstance(x, str) and x.startswith("cuda"))
+        or (isinstance(x, int) and x == 0)
+    )
+
+
+def _spaces_safe_to(self, *args, **kwargs):
+    """ZeroGPU emulates bare .to('cuda'), but LTX-2 uses non_blocking/copy."""
+    if args and _is_cuda_target(args[0]):
+        new_args = ("cuda",) + args[1:]
+        new_kwargs = {k: v for k, v in kwargs.items() if k not in ("non_blocking", "copy")}
+        return _original_tensor_to(self, *new_args, **new_kwargs)
+
+    if kwargs.get("device") is not None and _is_cuda_target(kwargs["device"]):
+        new_kwargs = {k: v for k, v in kwargs.items() if k not in ("non_blocking", "copy")}
+        new_kwargs["device"] = "cuda"
+        return _original_tensor_to(self, *args, **new_kwargs)
+
+    return _original_tensor_to(self, *args, **kwargs)
+
+
+torch.Tensor.to = _spaces_safe_to
+
+import gradio as gr
+import numpy as np
+from huggingface_hub import hf_hub_download, snapshot_download
+from safetensors import safe_open
+import requests
+
+from ltx_core.components.diffusion_steps import EulerDiffusionStep
+from ltx_core.components.noisers import GaussianNoiser
+from ltx_core.model.audio_vae import encode_audio as vae_encode_audio
+from ltx_core.model.upsampler import upsample_video
+from ltx_core.model.video_vae import TilingConfig, get_video_chunks_number, decode_video as vae_decode_video
+from ltx_core.quantization import QuantizationPolicy
+from ltx_core.types import Audio, AudioLatentShape, VideoPixelShape
+from ltx_pipelines.distilled import DistilledPipeline
+from ltx_pipelines.utils import euler_denoising_loop
+from ltx_pipelines.utils.args import ImageConditioningInput
+from ltx_pipelines.utils.constants import DISTILLED_SIGMA_VALUES, STAGE_2_DISTILLED_SIGMA_VALUES
+from ltx_pipelines.utils.helpers import (
+    cleanup_memory,
+    combined_image_conditionings,
+    denoise_video_only,
+    encode_prompts,
+    simple_denoising_func,
+)
+from ltx_pipelines.utils.media_io import decode_audio_from_file, encode_video
+from ltx_core.loader.primitives import LoraPathStrengthAndSDOps
+from ltx_core.loader.sd_ops import LTXV_LORA_COMFY_RENAMING_MAP
+
+logging.getLogger().setLevel(logging.INFO)
+
+MAX_SEED = np.iinfo(np.int32).max
+DEFAULT_VIDEO_PROMPT = (
+    "An astronaut hatches from a fragile egg on the surface of the Moon, "
+    "the shell cracking and peeling apart in gentle low-gravity motion. "
+    "Fine lunar dust lifts and drifts outward with each movement, floating "
+    "in slow arcs before settling back onto the ground."
+)
+DEFAULT_AUDIO_PROMPT = "Ambient space sounds with subtle crackling and distant echoes"
+DEFAULT_FRAME_RATE = 24.0
+
+RESOLUTIONS = {
+    "low":  {"16:9": (768, 512), "9:16": (512, 768), "1:1": (768, 768),
+             "4:3": (768, 576), "3:4": (576, 768), "21:9": (768, 384)},
+    "high": {"16:9": (1536, 1024), "9:16": (1024, 1536), "1:1": (1024, 1024),
+             "4:3": (1536, 1152), "3:4": (1152, 1536), "21:9": (1536, 768)},
+}
+
+
+class LTX23DistilledA2VPipeline(DistilledPipeline):
+    def __call__(
+        self,
+        video_prompt: str,
+        audio_prompt: str,
+        seed: int,
+        height: int,
+        width: int,
+        num_frames: int,
+        frame_rate: float,
+        images: list[ImageConditioningInput],
+        audio_path: str | None = None,
+        tiling_config: TilingConfig | None = None,
+        enhance_prompt: bool = False,
+    ):
+        print(f"Video prompt: {video_prompt}")
+        print(f"Audio prompt: {audio_prompt}")
+        if audio_path is None:
+            # Handle case when no audio is provided
+            raise NotImplementedError("This implementation requires audio input for audio-video generation")
+        
+        generator = torch.Generator(device=self.device).manual_seed(seed)
+        noiser = GaussianNoiser(generator=generator)
+        stepper = EulerDiffusionStep()
+        dtype = torch.bfloat16
+
+        # Encode video and audio prompts separately
+        (video_ctx,) = encode_prompts(
+            [video_prompt],
+            self.model_ledger,
+            enhance_first_prompt=enhance_prompt,
+            enhance_prompt_image=images[0].path if len(images) > 0 else None,
+        )
+        (audio_ctx,) = encode_prompts(
+            [audio_prompt],
+            self.model_ledger,
+            enhance_first_prompt=enhance_prompt,
+            enhance_prompt_image=images[0].path if len(images) > 0 else None,
+        )
+        video_context, audio_context = video_ctx.video_encoding, audio_ctx.audio_encoding
+
+        video_duration = num_frames / frame_rate
+        decoded_audio = decode_audio_from_file(audio_path, self.device, 0.0, video_duration)
+        if decoded_audio is None:
+            raise ValueError(f"Could not extract audio stream from {audio_path}")
+
+        encoded_audio_latent = vae_encode_audio(decoded_audio, self.model_ledger.audio_encoder())
+        audio_shape = AudioLatentShape.from_duration(batch=1, duration=video_duration, channels=8, mel_bins=16)
+        expected_frames = audio_shape.frames
+        actual_frames = encoded_audio_latent.shape[2]
+
+        if actual_frames > expected_frames:
+            encoded_audio_latent = encoded_audio_latent[:, :, :expected_frames, :]
+        elif actual_frames < expected_frames:
+            pad = torch.zeros(
+                encoded_audio_latent.shape[0],
+                encoded_audio_latent.shape[1],
+                expected_frames - actual_frames,
+                encoded_audio_latent.shape[3],
+                device=encoded_audio_latent.device,
+                dtype=encoded_audio_latent.dtype,
+            )
+            encoded_audio_latent = torch.cat([encoded_audio_latent, pad], dim=2)
+
+        video_encoder = self.model_ledger.video_encoder()
+        transformer = self.model_ledger.transformer()
+        stage_1_sigmas = torch.tensor(DISTILLED_SIGMA_VALUES, device=self.device)
+
+        def denoising_loop(sigmas, video_state, audio_state, stepper):
+            return euler_denoising_loop(
+                sigmas=sigmas,
+                video_state=video_state,
+                audio_state=audio_state,
+                stepper=stepper,
+                denoise_fn=simple_denoising_func(
+                    video_context=video_context,
+                    audio_context=audio_context,
+                    transformer=transformer,
+                ),
+            )
+
+        stage_1_output_shape = VideoPixelShape(
+            batch=1,
+            frames=num_frames,
+            width=width // 2,
+            height=height // 2,
+            fps=frame_rate,
+        )
+        stage_1_conditionings = combined_image_conditionings(
+            images=images,
+            height=stage_1_output_shape.height,
+            width=stage_1_output_shape.width,
+            video_encoder=video_encoder,
+            dtype=dtype,
+            device=self.device,
+        )
+        video_state = denoise_video_only(
+            output_shape=stage_1_output_shape,
+            conditionings=stage_1_conditionings,
+            noiser=noiser,
+            sigmas=stage_1_sigmas,
+            stepper=stepper,
+            denoising_loop_fn=denoising_loop,
+            components=self.pipeline_components,
+            dtype=dtype,
+            device=self.device,
+            initial_audio_latent=encoded_audio_latent,
+        )
+
+        torch.cuda.synchronize()
+        cleanup_memory()
+
+        upscaled_video_latent = upsample_video(
+            latent=video_state.latent[:1],
+            video_encoder=video_encoder,
+            upsampler=self.model_ledger.spatial_upsampler(),
+        )
+        stage_2_sigmas = torch.tensor(STAGE_2_DISTILLED_SIGMA_VALUES, device=self.device)
+        stage_2_output_shape = VideoPixelShape(batch=1, frames=num_frames, width=width, height=height, fps=frame_rate)
+        stage_2_conditionings = combined_image_conditionings(
+            images=images,
+            height=stage_2_output_shape.height,
+            width=stage_2_output_shape.width,
+            video_encoder=video_encoder,
+            dtype=dtype,
+            device=self.device,
+        )
+        video_state = denoise_video_only(
+            output_shape=stage_2_output_shape,
+            conditionings=stage_2_conditionings,
+            noiser=noiser,
+            sigmas=stage_2_sigmas,
+            stepper=stepper,
+            denoising_loop_fn=denoising_loop,
+            components=self.pipeline_components,
+            dtype=dtype,
+            device=self.device,
+            noise_scale=stage_2_sigmas[0],
+            initial_video_latent=upscaled_video_latent,
+            initial_audio_latent=encoded_audio_latent,
+        )
+
+        torch.cuda.synchronize()
+        del transformer
+        del video_encoder
+        cleanup_memory()
+
+        decoded_video = vae_decode_video(
+            video_state.latent,
+            self.model_ledger.video_decoder(),
+            tiling_config,
+            generator,
+        )
+        original_audio = Audio(
+            waveform=decoded_audio.waveform.squeeze(0),
+            sampling_rate=decoded_audio.sampling_rate,
+        )
+        return decoded_video, original_audio
+
+
+# Model repos
+LTX_MODEL_REPO = "Lightricks/LTX-2.3"
+GEMMA_REPO ="Lightricks/gemma-3-12b-it-qat-q4_0-unquantized"
+
+print("=" * 80)
+print("Downloading LTX-2.3 distilled model + Gemma...")
+print("=" * 80)
+
+_legacy_lora_cache_dir = Path("lora_cache")
+if _legacy_lora_cache_dir.exists():
+    shutil.rmtree(_legacy_lora_cache_dir, ignore_errors=True)
+
+current_lora_key: str | None = None
+PENDING_LORA_KEY: str | None = None
+PENDING_LORA_STATE: dict[str, torch.Tensor] | None = None
+PENDING_LORA_STATUS: str = "No LoRA state prepared yet."
+
+weights_dir = Path("weights")
+weights_dir.mkdir(exist_ok=True)
+checkpoint_path = hf_hub_download(
+    repo_id="TenStrip/LTX2.3-10Eros",
+    filename="10Eros_v1.3_bf16.safetensors",
+    local_dir=str(weights_dir),
+    local_dir_use_symlinks=False,
+)
+
+spatial_upsampler_path = hf_hub_download(repo_id=LTX_MODEL_REPO, filename="ltx-2.3-spatial-upscaler-x2-1.1.safetensors")
+gemma_root = snapshot_download(repo_id=GEMMA_REPO)
+
+LORA_REPO = "dagloop5/LoRA"
+print("=" * 80)
+print("Downloading LoRA adapters from dagloop5/LoRA...")
+print("=" * 80)
+singularity_lora_path = hf_hub_download(repo_id="TenStrip/LTX2.3_Distilled_Lora_1.1_Experiments", filename="ltx-2.3-22b-distilled-lora-1.1_rank72_energy.safetensors")
+teneros_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="LTX2.3-Furry-2D-NSFW-Multi-Purpose-Lora+Cum.safetensors")
+sulphur_lora_path =hf_hub_download(repo_id=LORA_REPO, filename="ltx23E28093SlowMotion26.Pkrs.safetensors")
+pose_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="LTX2_3_NSFW_furry_concat_v2.safetensors")
+general_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="LTX2.3_reasoning_Sulphur-2_I2V_V4.safetensors")
+motion_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="Sulphur_LTX 2.3_better _NSFW_motion.safetensors")
+dreamlay_lora_path = hf_hub_download(repo_id="lynaNSFW/DR34ML4Y_AIO_NSFW_LTX23", filename="DR34ML4Y_LTXXX_V2.safetensors") 
+mself_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="LTX2.3_2d_NSFW_motion_enhancer.safetensors") 
+dramatic_lora_path = hf_hub_download(repo_id="Muapi/valiantcat-ltx-2.3-transition-lora", filename="valiantcat-ltx-2.3-transition-lora.safetensors") 
+fluid_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="Cr3ampi3_animation_sulphur-2_i2v_v1.0.safetensors") 
+liquid_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="liquid_wet_dr1pp_ltx2_v1.0_scaled.safetensors") 
+demopose_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="clapping-cheeks-audio-v001-alpha.safetensors")
+voice_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="hentai_voice_ltx23_v2.comfy.safetensors")
+realism_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="FurryenhancerLTX2.3V4.094fused.safetensors")
+transition_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="LTX_10Eros_takerpov_V2.safetensors") 
+physics_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="LTX2.3_Physics_V2_000002000.safetensors")
+reasoning_lora_path = hf_hub_download(repo_id="LiconStudio/Ltx2.3-VBVR-lora-I2V", filename="Ltx2.3-Licon-VBVR-I2V-390K-R32.safetensors")
+twostep_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="LTX2.3_Multi_step_video_reasoning_V0.1.safetensors")
+mcfurry_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="mvmt_lora_v2_600.safetensors")
+dm_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="Doggy_mission_sulphur-2_v0.5.safetensors")
+praxis_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="Penile_Praxis_V4.safetensors")
+threed_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="ltx2-3d-animations-12500-steps-k3nk.safetensors")
+concept_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="ltx23_nsfw_helper_multi_concept_lora_v2.safetensors")
+bulge_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="stomach_bulge_10eros_sulphur_v1.safetensors")
+
+pipeline = LTX23DistilledA2VPipeline(
+    distilled_checkpoint_path=checkpoint_path,
+    spatial_upsampler_path=spatial_upsampler_path,
+    gemma_root=gemma_root,
+    loras=[],
+    quantization=QuantizationPolicy.fp8_cast(),
+)
+
+def _make_lora_key(singularity_strength, teneros_strength, sulphur_strength, pose_strength, general_strength, motion_strength, dreamlay_strength, mself_strength, dramatic_strength, fluid_strength, liquid_strength, demopose_strength, voice_strength, realism_strength, transition_strength, physics_strength, reasoning_strength, twostep_strength, mcfurry_strength, dm_strength, praxis_strength, threed_strength, concept_strength, bulge_strength) -> tuple[str, str]:
+    rx, ra, rb, rp, rg, rm, rd, rs, rr, rf, rl, ro, rv, re, rt, ry, ri, rw, mc, dm, pr, td, co, bu = [round(float(x), 2) for x in [singularity_strength, teneros_strength, sulphur_strength, pose_strength, general_strength, motion_strength, dreamlay_strength, mself_strength, dramatic_strength, fluid_strength, liquid_strength, demopose_strength, voice_strength, realism_strength, transition_strength, physics_strength, reasoning_strength, twostep_strength, mcfurry_strength, dm_strength, praxis_strength, threed_strength, concept_strength, bulge_strength]]
+    key_str = f"{singularity_lora_path}:{rx}|{teneros_lora_path}:{ra}|{sulphur_lora_path}:{rb}|{pose_lora_path}:{rp}|{general_lora_path}:{rg}|{motion_lora_path}:{rm}|{dreamlay_lora_path}:{rd}|{mself_lora_path}:{rs}|{dramatic_lora_path}:{rr}|{fluid_lora_path}:{rf}|{liquid_lora_path}:{rl}|{demopose_lora_path}:{ro}|{voice_lora_path}:{rv}|{realism_lora_path}:{re}|{transition_lora_path}:{rt}|{physics_lora_path}:{ry}|{reasoning_lora_path}:{ri}|{twostep_lora_path}:{rw}|{mcfurry_lora_path}:{mc}|{dm_lora_path}:{dm}|{praxis_lora_path}:{pr}|{threed_lora_path}:{td}|{concept_lora_path}:{co}|{bulge_lora_path}:{bu}"
+    key = hashlib.sha256(key_str.encode("utf-8")).hexdigest()
+    return key, key_str
+
+def prepare_lora_cache(
+    singularity_strength, teneros_strength, sulphur_strength, pose_strength, general_strength, motion_strength, dreamlay_strength, mself_strength, dramatic_strength, fluid_strength, liquid_strength, demopose_strength, voice_strength, realism_strength, transition_strength, physics_strength, reasoning_strength, twostep_strength, mcfurry_strength, dm_strength, praxis_strength, threed_strength, concept_strength, bulge_strength,
+    progress=gr.Progress(track_tqdm=True),
+):
+    global PENDING_LORA_KEY, PENDING_LORA_STATE, PENDING_LORA_STATUS
+    ledger = pipeline.model_ledger
+    key, _ = _make_lora_key(singularity_strength, teneros_strength, sulphur_strength, pose_strength, general_strength, motion_strength, dreamlay_strength, mself_strength, dramatic_strength, fluid_strength, liquid_strength, demopose_strength, voice_strength, realism_strength, transition_strength, physics_strength, reasoning_strength, twostep_strength, mcfurry_strength, dm_strength, praxis_strength, threed_strength, concept_strength, bulge_strength)
+    progress(0.05, desc="Preparing LoRA state")
+    entries = [
+        (singularity_lora_path, round(float(singularity_strength), 2)), (teneros_lora_path, round(float(teneros_strength), 2)), (sulphur_lora_path, round(float(sulphur_strength), 2)), (pose_lora_path, round(float(pose_strength), 2)), (general_lora_path, round(float(general_strength), 2)), (motion_lora_path, round(float(motion_strength), 2)), (dreamlay_lora_path, round(float(dreamlay_strength), 2)), (mself_lora_path, round(float(mself_strength), 2)), (dramatic_lora_path, round(float(dramatic_strength), 2)), (fluid_lora_path, round(float(fluid_strength), 2)), (liquid_lora_path, round(float(liquid_strength), 2)), (demopose_lora_path, round(float(demopose_strength), 2)), (voice_lora_path, round(float(voice_strength), 2)), (realism_lora_path, round(float(realism_strength), 2)), (transition_lora_path, round(float(transition_strength), 2)), (physics_lora_path, round(float(physics_strength), 2)), (reasoning_lora_path, round(float(reasoning_strength), 2)), (twostep_lora_path, round(float(twostep_strength), 2)), (mcfurry_lora_path, round(float(mcfurry_strength), 2)), (dm_lora_path, round(float(dm_strength), 2)), (praxis_lora_path, round(float(praxis_strength), 2)), (threed_lora_path, round(float(threed_strength), 2)), (concept_lora_path, round(float(concept_strength), 2)), (bulge_lora_path, round(float(bulge_strength), 2)),
+    ]
+    loras_for_builder = [LoraPathStrengthAndSDOps(path, strength, LTXV_LORA_COMFY_RENAMING_MAP) for path, strength in entries if path is not None and float(strength) != 0.0]
+    if not loras_for_builder:
+        PENDING_LORA_KEY = None
+        PENDING_LORA_STATE = None
+        PENDING_LORA_STATUS = "No non-zero LoRA strengths selected; nothing to prepare."
+        return PENDING_LORA_STATUS
+    try:
+        progress(0.35, desc="Building fused CPU transformer")
+        tmp_ledger = pipeline.model_ledger.__class__(dtype=ledger.dtype, device=torch.device("cpu"), checkpoint_path=str(checkpoint_path), spatial_upsampler_path=str(spatial_upsampler_path), gemma_root_path=str(gemma_root), loras=tuple(loras_for_builder), quantization=QuantizationPolicy.fp8_cast())
+        new_transformer_cpu = tmp_ledger.transformer()
+        progress(0.70, desc="Extracting fused state_dict")
+        state = {k: v.detach().cpu().contiguous() for k, v in new_transformer_cpu.state_dict().items()}
+        PENDING_LORA_KEY = key
+        PENDING_LORA_STATE = state
+        PENDING_LORA_STATUS = "Built LoRA state (ready to apply)."
+        return PENDING_LORA_STATUS
+    except Exception as e:
+        PENDING_LORA_KEY = None
+        PENDING_LORA_STATE = None
+        PENDING_LORA_STATUS = f"LoRA prepare failed: {type(e).__name__}: {e}"
+        return PENDING_LORA_STATUS
+    finally:
+        gc.collect()
+
+def apply_prepared_lora_state_to_pipeline():
+    global current_lora_key, PENDING_LORA_KEY, PENDING_LORA_STATE, PENDING_LORA_STATUS
+    if PENDING_LORA_KEY is None: return False
+    if current_lora_key == PENDING_LORA_KEY:
+        if PENDING_LORA_STATE is not None: PENDING_LORA_STATE = None
+        return True
+    if PENDING_LORA_STATE is None: return False
+    with torch.no_grad():
+        _transformer.load_state_dict(PENDING_LORA_STATE, strict=False)
+    current_lora_key = PENDING_LORA_KEY
+    PENDING_LORA_STATE = None
+    PENDING_LORA_STATUS = "LoRA state applied to pipeline."
+    return True
+
+print("Preloading all models...")
+ledger = pipeline.model_ledger
+_transformer = ledger.transformer()
+_video_encoder = ledger.video_encoder()
+_video_decoder = ledger.video_decoder()
+_audio_encoder = ledger.audio_encoder()
+_audio_decoder = ledger.audio_decoder()
+_vocoder = ledger.vocoder()
+_spatial_upsampler = ledger.spatial_upsampler()
+_text_encoder = ledger.text_encoder()
+_embeddings_processor = ledger.gemma_embeddings_processor()
+
+ledger.transformer = lambda: _transformer
+ledger.video_encoder = lambda: _video_encoder
+ledger.video_decoder = lambda: _video_decoder
+ledger.audio_encoder = lambda: _audio_encoder
+ledger.audio_decoder = lambda: _audio_decoder
+ledger.vocoder = lambda: _vocoder
+ledger.spatial_upsampler = lambda: _spatial_upsampler
+ledger.text_encoder = lambda: _text_encoder
+ledger.gemma_embeddings_processor = lambda: _embeddings_processor
+print("All models preloaded!")
+
+def log_memory(tag: str):
+    if torch.cuda.is_available():
+        allocated = torch.cuda.memory_allocated() / 1024**3
+        print(f"[VRAM {tag}] allocated={allocated:.2f}GB")
+
+def detect_aspect_ratio(image) -> str:
+    if image is None: return "16:9"
+    w, h = (image.size if hasattr(image, "size") else image.shape[:2][::-1])
+    ratio = w / h
+    candidates = {"16:9": 16 / 9, "9:16": 9 / 16, "1:1": 1.0}
+    return min(candidates, key=lambda k: abs(ratio - candidates[k]))
+
+def on_image_upload(first_image, last_image, high_res):
+    ref_image = first_image if first_image is not None else last_image
+    aspect = detect_aspect_ratio(ref_image)
+    tier = "high" if high_res else "low"
+    w, h = RESOLUTIONS[tier][aspect]
+    return gr.update(value=w), gr.update(value=h)
+
+def on_highres_toggle(first_image, last_image, high_res):
+    ref_image = first_image if first_image is not None else last_image
+    aspect = detect_aspect_ratio(ref_image)
+    tier = "high" if high_res else "low"
+    w, h = RESOLUTIONS[tier][aspect]
+    return gr.update(value=w), gr.update(value=h)
+
+def get_gpu_duration(first_image, last_image, input_audio, video_prompt, audio_prompt, duration, gpu_duration, enhance_prompt, seed, randomize_seed, height, width, singularity_strength, teneros_strength, sulphur_strength, pose_strength, general_strength, motion_strength, dreamlay_strength, mself_strength, dramatic_strength, fluid_strength, liquid_strength, demopose_strength, voice_strength, realism_strength, transition_strength, physics_strength, reasoning_strength, twostep_strength, mcfurry_strength, dm_strength, praxis_strength, threed_strength, concept_strength, bulge_strength, progress=None):
+    return int(gpu_duration)
+
+@spaces.GPU(size="xlarge", duration=get_gpu_duration)
+@torch.inference_mode()
+def generate_video(first_image, last_image, input_audio, video_prompt, audio_prompt, duration, gpu_duration, enhance_prompt, seed, randomize_seed, height, width, singularity_strength, teneros_strength, sulphur_strength, pose_strength, general_strength, motion_strength, dreamlay_strength, mself_strength, dramatic_strength, fluid_strength, liquid_strength, demopose_strength, voice_strength, realism_strength, transition_strength, physics_strength, reasoning_strength, twostep_strength, mcfurry_strength, dm_strength, praxis_strength, threed_strength, concept_strength, bulge_strength, progress=gr.Progress(track_tqdm=True)):
+    try:
+        current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
+        frame_rate = DEFAULT_FRAME_RATE
+        num_frames = int(duration * frame_rate) + 1
+        num_frames = ((num_frames - 1 + 7) // 8) * 8 + 1
+        images = []
+        output_dir = Path("outputs")
+        output_dir.mkdir(exist_ok=True)
+        if first_image is not None:
+            temp_first_path = output_dir / f"temp_first_{current_seed}.jpg"
+            if hasattr(first_image, "save"): first_image.save(temp_first_path)
+            else: temp_first_path = Path(first_image)
+            images.append(ImageConditioningInput(path=str(temp_first_path), frame_idx=0, strength=1.0))
+        if last_image is not None:
+            temp_last_path = output_dir / f"temp_last_{current_seed}.jpg"
+            if hasattr(last_image, "save"): last_image.save(temp_last_path)
+            else: temp_last_path = Path(last_image)
+            images.append(ImageConditioningInput(path=str(temp_last_path), frame_idx=num_frames - 1, strength=1.0))
+        tiling_config = TilingConfig.default()
+        video_chunks_number = get_video_chunks_number(num_frames, tiling_config)
+        apply_prepared_lora_state_to_pipeline()
+        video, audio = pipeline(
+            video_prompt=video_prompt, 
+            audio_prompt=audio_prompt, 
+            seed=current_seed, 
+            height=int(height), 
+            width=int(width), 
+            num_frames=num_frames, 
+            frame_rate=frame_rate, 
+            images=images, 
+            audio_path=input_audio, 
+            tiling_config=tiling_config, 
+            enhance_prompt=enhance_prompt
+        )
+        output_path = tempfile.mktemp(suffix=".mp4")
+        encode_video(video=video, fps=frame_rate, audio=audio, output_path=output_path, video_chunks_number=video_chunks_number)
+        return str(output_path), current_seed
+    except Exception as e:
+        return None, current_seed
+
+with gr.Blocks(title="LTX-2.3 Distilled") as demo:
+    gr.Markdown("# LTX-2.3 F2LF with Fast Audio-Video Generation with Frame Conditioning")
+    with gr.Row():
+        with gr.Column():
+            with gr.Row():
+                first_image = gr.Image(label="First Frame (Optional)", type="pil")
+                last_image = gr.Image(label="Last Frame (Optional)", type="pil")
+            input_audio = gr.Audio(label="Audio Input (Optional)", type="filepath")
+            
+            with gr.Tab("Prompts"):
+                video_prompt = gr.Textbox(
+                    label="Video Prompt", 
+                    value=DEFAULT_VIDEO_PROMPT, 
+                    lines=3
+                )
+                audio_prompt = gr.Textbox(
+                    label="Audio Prompt", 
+                    value=DEFAULT_AUDIO_PROMPT, 
+                    lines=3
+                )
+                
+            duration = gr.Slider(label="Duration (seconds)", minimum=1.0, maximum=30.0, value=10.0, step=0.1)
+            generate_btn = gr.Button("Generate Video", variant="primary", size="lg")
+            with gr.Accordion("Advanced Settings", open=False):
+                seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, value=10, step=1)
+                randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
+                with gr.Row():
+                    width = gr.Number(label="Width", value=1536, precision=0)
+                    height = gr.Number(label="Height", value=1024, precision=0)
+                with gr.Row():
+                    enhance_prompt = gr.Checkbox(label="Enhance Prompt", value=False)
+                    high_res = gr.Checkbox(label="High Resolution", value=True)
+                with gr.Column():
+                    gr.Markdown("### LoRA adapter strengths")
+                    singularity_strength = gr.Slider(label="Distilled Lora strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    teneros_strength = gr.Slider(label="Multipurpose furry strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    sulphur_strength = gr.Slider(label="Floaty/Slow Motion Reducer strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    pose_strength = gr.Slider(label="Anthro Enhancer strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    general_strength = gr.Slider(label="Reasoning Enhancer strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    motion_strength = gr.Slider(label="Anthro Posing Helper strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    dreamlay_strength = gr.Slider(label="Dreamlay strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    mself_strength = gr.Slider(label="2D enhancer strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    dramatic_strength = gr.Slider(label="Transition enhancer strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    fluid_strength = gr.Slider(label="Fluid Helper strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    liquid_strength = gr.Slider(label="Liquid Helper strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    demopose_strength = gr.Slider(label="Audio Helper strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    voice_strength = gr.Slider(label="Voice Helper strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    realism_strength = gr.Slider(label="Anthro Realism strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    transition_strength = gr.Slider(label="POV strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    physics_strength = gr.Slider(label="Physics strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    reasoning_strength = gr.Slider(label="Official Reasoning strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    twostep_strength = gr.Slider(label="Two Step Reasoning strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    mcfurry_strength = gr.Slider(label="I2V Motion enhancer strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    dm_strength = gr.Slider(label="DM3D strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    praxis_strength = gr.Slider(label="Praxis strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    threed_strength = gr.Slider(label="3D animation strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    concept_strength = gr.Slider(label="Conceptual strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                    bulge_strength = gr.Slider(label="Bulge strength", minimum=0.0, maximum=2.0, value=0.0, step=0.01)
+                prepare_lora_btn = gr.Button("Prepare / Load LoRA Cache", variant="secondary")
+                lora_status = gr.Textbox(label="LoRA Cache Status", value="No LoRA state prepared yet.", interactive=False)
+        with gr.Column():
+            output_video = gr.Video(label="Generated Video", autoplay=False)
+            gpu_duration = gr.Slider(label="ZeroGPU duration (seconds)", minimum=30.0, maximum=240.0, value=75.0, step=1.0)
+
+    first_image.change(fn=on_image_upload, inputs=[first_image, last_image, high_res], outputs=[width, height])
+    last_image.change(fn=on_image_upload, inputs=[first_image, last_image, high_res], outputs=[width, height])
+    high_res.change(fn=on_highres_toggle, inputs=[first_image, last_image, high_res], outputs=[width, height])
+    prepare_lora_btn.click(fn=prepare_lora_cache, inputs=[singularity_strength, teneros_strength, sulphur_strength, pose_strength, general_strength, motion_strength, dreamlay_strength, mself_strength, dramatic_strength, fluid_strength, liquid_strength, demopose_strength, voice_strength, realism_strength, transition_strength, physics_strength, reasoning_strength, twostep_strength, mcfurry_strength, dm_strength, praxis_strength, threed_strength, concept_strength, bulge_strength], outputs=[lora_status])
+    generate_btn.click(fn=generate_video, inputs=[first_image, last_image, input_audio, video_prompt, audio_prompt, duration, gpu_duration, enhance_prompt, seed, randomize_seed, height, width, singularity_strength, teneros_strength, sulphur_strength, pose_strength, general_strength, motion_strength, dreamlay_strength, mself_strength, dramatic_strength, fluid_strength, liquid_strength, demopose_strength, voice_strength, realism_strength, transition_strength, physics_strength, reasoning_strength, twostep_strength, mcfurry_strength, dm_strength, praxis_strength, threed_strength, concept_strength, bulge_strength], outputs=[output_video, seed])
+
+if __name__ == "__main__":
+    demo.launch(theme=gr.themes.Citrus(), css=".fillable{max-width: 1200px !important}")