Update app.py

app.py

@@ -106,8 +106,36 @@ RESOLUTIONS = {
 }
 
 
-class LTX23DistilledA2VPipeline(DistilledPipeline):
-    """DistilledPipeline with optional audio conditioning."""
+class LTX23DistilledA2VPipeline:
+
+    def __init__(
+        self,
+        distilled_checkpoint_path: str,
+        spatial_upsampler_path: str,
+        gemma_root: str,
+        loras: tuple,
+        quantization: QuantizationPolicy | None = None,
+    ):
+        from ltx_pipelines.utils import ModelLedger, denoise_audio_video
+        from ltx_pipelines.utils.types import PipelineComponents
+
+        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        self.dtype = torch.bfloat16
+
+        self.model_ledger = ModelLedger(
+            dtype=self.dtype,
+            device=self.device,
+            checkpoint_path=distilled_checkpoint_path,
+            gemma_root_path=gemma_root,
+            spatial_upsampler_path=spatial_upsampler_path,
+            loras=loras,
+            quantization=quantization,
+        )
+
+        self.pipeline_components = PipelineComponents(
+            dtype=self.dtype,
+            device=self.device,
+        )
 
     def __call__(
         self,
@@ -118,24 +146,10 @@ class LTX23DistilledA2VPipeline(DistilledPipeline):
         num_frames: int,
         frame_rate: float,
         images: list[ImageConditioningInput],
-        audio_path: str | None = None,
         tiling_config: TilingConfig | None = None,
         enhance_prompt: bool = False,
     ):
-        # Standard path when no audio input is provided.
         print(prompt)
-        if audio_path is None:
-            return super().__call__(
-                prompt=prompt,
-                seed=seed,
-                height=height,
-                width=width,
-                num_frames=num_frames,
-                frame_rate=frame_rate,
-                images=images,
-                tiling_config=tiling_config,
-                enhance_prompt=enhance_prompt,
-            )
 
         generator = torch.Generator(device=self.device).manual_seed(seed)
         noiser = GaussianNoiser(generator=generator)
@@ -151,9 +165,6 @@ class LTX23DistilledA2VPipeline(DistilledPipeline):
         video_context, audio_context = ctx_p.video_encoding, ctx_p.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)
@@ -262,18 +273,17 @@ class LTX23DistilledA2VPipeline(DistilledPipeline):
             tiling_config,
             generator,
         )
-        original_audio = Audio(
-            waveform=decoded_audio.waveform.squeeze(0),
-            sampling_rate=decoded_audio.sampling_rate,
+        decoded_audio_output = vae_decode_audio(
+            audio_state.latent,
+            self.model_ledger.audio_decoder(),
+            self.model_ledger.vocoder(),
         )
-        return decoded_video, original_audio
+        return decoded_video, decoded_audio_output
 
 
 # Model repos
 LTX_MODEL_REPO = "Lightricks/LTX-2.3"
 GEMMA_REPO ="Lightricks/gemma-3-12b-it-qat-q4_0-unquantized"
-GEMMA_ABLITERATED_REPO = "Sikaworld1990/gemma-3-12b-it-abliterated-sikaworld-high-fidelity-edition-Ltx-2"
-GEMMA_ABLITERATED_FILE = "gemma-3-12b-it-abliterated-sikaworld-high-fidelity-edition.safetensors"
 
 # Download model checkpoints
 print("=" * 80)
@@ -297,73 +307,6 @@ checkpoint_path = hf_hub_download(
     local_dir=str(weights_dir),
     local_dir_use_symlinks=False,
 )
-
-print("[Gemma] Setting up abliterated Gemma text encoder...")
-MERGED_WEIGHTS = "/tmp/abliterated_gemma_merged.safetensors"
-gemma_root = "/tmp/abliterated_gemma"
-os.makedirs(gemma_root, exist_ok=True)
-
-gemma_official_dir = snapshot_download(
-    repo_id=GEMMA_REPO,
-    ignore_patterns=["*.safetensors", "*.safetensors.index.json"],
-)
-
-for fname in os.listdir(gemma_official_dir):
-    src = os.path.join(gemma_official_dir, fname)
-    dst = os.path.join(gemma_root, fname)
-    if os.path.isfile(src) and not fname.endswith(".safetensors") and fname != "model.safetensors.index.json":
-        if not os.path.exists(dst):
-            os.symlink(src, dst)
-
-if os.path.exists(MERGED_WEIGHTS):
-    print("[Gemma] Using cached merged weights")
-else:
-    abliterated_weights_path = hf_hub_download(
-        repo_id=GEMMA_ABLITERATED_REPO,
-        filename=GEMMA_ABLITERATED_FILE,
-    )
-    index_path = hf_hub_download(
-        repo_id=GEMMA_REPO,
-        filename="model.safetensors.index.json"
-    )
-    with open(index_path) as f:
-        weight_index = json.load(f)
-
-    vision_keys = {}
-    for key, shard in weight_index["weight_map"].items():
-        if "vision_tower" in key or "multi_modal_projector" in key:
-            vision_keys[key] = shard
-    needed_shards = set(vision_keys.values())
-
-    shard_paths = {}
-    for shard_name in needed_shards:
-        shard_paths[shard_name] = hf_hub_download(
-            repo_id=GEMMA_REPO,
-            filename=shard_name
-        )
-
-    _fp8_types = {torch.float8_e4m3fn, torch.float8_e5m2}
-    raw = load_file(abliterated_weights_path)
-    merged = {}
-    for key, tensor in raw.items():
-        t = tensor.to(torch.bfloat16) if tensor.dtype in _fp8_types else tensor
-        merged[f"language_model.{key}"] = t
-    del raw
-
-    for key, shard_name in vision_keys.items():
-        with safe_open(shard_paths[shard_name], framework="pt") as f:
-            merged[key] = f.get_tensor(key)
-
-    save_file(merged, MERGED_WEIGHTS)
-    del merged
-    gc.collect()
-
-weight_link = os.path.join(gemma_root, "model.safetensors")
-if os.path.exists(weight_link):
-    os.remove(weight_link)
-os.symlink(MERGED_WEIGHTS, weight_link)
-print(f"[Gemma] Root ready: {gemma_root}")
-
 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)
 
@@ -381,7 +324,7 @@ motion_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="motion_helper.sa
 dreamlay_lora_path = hf_hub_download(repo_id="lynaNSFW/DR34ML4Y_AIO_NSFW_LTX23", filename="DR34ML4Y_LTXXX_V2.safetensors") # m15510n4ry, bl0wj0b, d0ubl3_bj, d0gg1e, c0wg1rl
 mself_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="Furry Hyper Masturbation - LTX-2 I2V v1.safetensors") # Hyperfap
 dramatic_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="LTX-2.3 - Orgasm.safetensors") # "[He | She] is having am orgasm." (am or an?)
-fluid_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="LTX2.3_CREAMPIE_ANIMATION-V0.1.safetensors") # cum
+fluid_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="Cr3ampi3_animation_sulphur-2_i2v_v1.0.safetensors") # cr3ampi3 animation
 liquid_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="liquid_wet_dr1pp_ltx2_v1.0_scaled.safetensors") # wet dr1pp
 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.safetensors")
@@ -597,9 +540,8 @@ ledger = pipeline.model_ledger
 # Save the original factory methods so we can rebuild individual components later.
 # These are bound callables on ledger that will call the builder when invoked.
 _orig_transformer_factory = ledger.transformer
-_orig_video_encoder_factory = ledger.video_encoder
+_orig_video_encoder_factory = ledger.video_encode
 _orig_video_decoder_factory = ledger.video_decoder
-_orig_audio_encoder_factory = ledger.audio_encoder
 _orig_audio_decoder_factory = ledger.audio_decoder
 _orig_vocoder_factory = ledger.vocoder
 _orig_spatial_upsampler_factory = ledger.spatial_upsampler
@@ -610,7 +552,6 @@ _orig_gemma_embeddings_factory = ledger.gemma_embeddings_processor
 _transformer = _orig_transformer_factory()
 _video_encoder = _orig_video_encoder_factory()
 _video_decoder = _orig_video_decoder_factory()
-_audio_encoder = _orig_audio_encoder_factory()
 _audio_decoder = _orig_audio_decoder_factory()
 _vocoder = _orig_vocoder_factory()
 _spatial_upsampler = _orig_spatial_upsampler_factory()
@@ -622,7 +563,6 @@ _embeddings_processor = _orig_gemma_embeddings_factory()
 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
@@ -678,7 +618,6 @@ def on_highres_toggle(first_image, last_image, high_res):
 def get_gpu_duration(
     first_image,
     last_image,
-    input_audio,
     prompt: str,
     duration: float = 0.0,
     gpu_duration: float = 0.0,
@@ -711,7 +650,6 @@ def get_gpu_duration(
 def generate_video(
     first_image,
     last_image,
-    input_audio,
     prompt: str,
     duration: float = 0.0,
     gpu_duration: float = 0.0,
@@ -784,7 +722,6 @@ def generate_video(
             num_frames=num_frames,
             frame_rate=frame_rate,
             images=images,
-            audio_path=input_audio,
             tiling_config=tiling_config,
             enhance_prompt=enhance_prompt,
         )
@@ -819,7 +756,6 @@ with gr.Blocks(title="LTX-2.3 Distilled") as demo:
             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")
             prompt = gr.Textbox(
                 label="Prompt",
                 info="for best results - make it as elaborate as possible",
@@ -920,49 +856,6 @@ with gr.Blocks(title="LTX-2.3 Distilled") as demo:
                 step=1.0,
             )
 
-    gr.Examples(
-        examples=[
-            [
-                None,
-                "pinkknit.jpg",
-                None,
-                "The camera falls downward through darkness as if dropped into a tunnel. "
-                "As it slows, five friends wearing pink knitted hats and sunglasses lean "
-                "over and look down toward the camera with curious expressions. The lens "
-                "has a strong fisheye effect, creating a circular frame around them. They "
-                "crowd together closely, forming a symmetrical cluster while staring "
-                "directly into the lens.",
-                3.0,
-                80.0,
-                False,
-                42,
-                True,
-                1024,
-                1024,
-                0.0,  # pose_strength (example)
-                0.0,  # general_strength (example)
-                0.0,  # motion_strength (example)
-                0.0,
-                0.0,
-                0.0,
-                0.0,
-                0.0,
-                0.0,
-                0.0,
-                0.0,
-                0.0,
-                0.0,
-                0.0,
-                0.0,
-            ],
-        ],
-        inputs=[
-            first_image, last_image, input_audio, prompt, duration, gpu_duration,
-            enhance_prompt, seed, randomize_seed, height, width,
-            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,
-        ],
-    )
-
     first_image.change(
         fn=on_image_upload,
         inputs=[first_image, last_image, high_res],
@@ -990,7 +883,7 @@ with gr.Blocks(title="LTX-2.3 Distilled") as demo:
     generate_btn.click(
         fn=generate_video,
         inputs=[
-            first_image, last_image, input_audio, prompt, duration, gpu_duration, enhance_prompt,
+            first_image, last_image, prompt, duration, gpu_duration, enhance_prompt,
             seed, randomize_seed, height, width,
             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,
         ],