Update app.py

app.py

@@ -267,18 +267,78 @@ checkpoint_path = hf_hub_download(repo_id=LTX_MODEL_REPO, filename="ltx-2.3-22b-
 spatial_upsampler_path = hf_hub_download(repo_id=LTX_MODEL_REPO, filename="ltx-2.3-spatial-upscaler-x2-1.0.safetensors")
 gemma_root = snapshot_download(repo_id=GEMMA_REPO)
 
+# ---- Insert block (LoRA downloads) between lines 268 and 269 ----
+# LoRA repo + download the requested LoRA adapters
+LORA_REPO = "dagloop5/LoRA"
+
+print("=" * 80)
+print("Downloading LoRA adapters from dagloop5/LoRA...")
+print("=" * 80)
+pose_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="pose_enhancer.safetensors")
+general_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="general_enhancer.safetensors")
+motion_lora_path = hf_hub_download(repo_id=LORA_REPO, filename="motion_helper.safetensors")
+
+print(f"Pose LoRA: {pose_lora_path}")
+print(f"General LoRA: {general_lora_path}")
+print(f"Motion LoRA: {motion_lora_path}")
+# ----------------------------------------------------------------
+
 print(f"Checkpoint: {checkpoint_path}")
 print(f"Spatial upsampler: {spatial_upsampler_path}")
 print(f"Gemma root: {gemma_root}")
 
 # Initialize pipeline WITH text encoder and optional audio support
+# ---- Replace block (pipeline init) lines 275-281 ----
 pipeline = LTX23DistilledA2VPipeline(
     distilled_checkpoint_path=checkpoint_path,
     spatial_upsampler_path=spatial_upsampler_path,
     gemma_root=gemma_root,
-    loras=[],
-    quantization=QuantizationPolicy.fp8_cast(),
+    # initial LoRAs (strengths set to zero by default; will be adjusted at runtime)
+    loras=[
+        (pose_lora_path, 0.0),
+        (general_lora_path, 0.0),
+        (motion_lora_path, 0.0),
+    ],
+    quantization=QuantizationPolicy.fp8_cast(),  # keep FP8 quantization unchanged
 )
+# ----------------------------------------------------------------
+
+# ---- Insert block: helper to apply LoRA strengths at runtime (between lines 281 and 283) ----
+def apply_loras_to_pipeline(pose_strength: float, general_strength: float, motion_strength: float):
+    """
+    Best-effort attempt to set LoRA strengths on the existing pipeline without creating a new pipeline object.
+    Tries common APIs that model_ledger / pipeline implementations may provide; falls back to setting
+    pipeline.loras (for lazy fusion) and logs a note.
+    """
+    # Build the spec list (keep ordering consistent with the pipeline loras list created above)
+    lora_specs = [
+        (pose_lora_path, float(pose_strength)),
+        (general_lora_path, float(general_strength)),
+        (motion_lora_path, float(motion_strength)),
+    ]
+    try:
+        # Preferred: model_ledger exposes an apply_loras API that fuses adapters in-place.
+        if hasattr(pipeline, "model_ledger") and hasattr(pipeline.model_ledger, "apply_loras"):
+            print("[LoRA] Applying LoRAs via pipeline.model_ledger.apply_loras(...)")
+            pipeline.model_ledger.apply_loras(lora_specs)
+            return
+
+        # Alternative: pipeline may have helper 'set_loras' or similar
+        if hasattr(pipeline, "set_loras"):
+            print("[LoRA] Applying LoRAs via pipeline.set_loras(...)")
+            pipeline.set_loras(lora_specs)
+            return
+
+        # Fallback: set an attribute that some pipeline implementations inspect lazily.
+        print("[LoRA] Setting pipeline.loras attribute (fallback). If pipeline supports lazy fusion, new strengths will be used.")
+        pipeline.loras = lora_specs
+        return
+
+    except Exception as e:
+        # Non-fatal: advise user and keep running with whatever pipeline state exists.
+        print(f"[LoRA] Warning: failed to apply LoRAs at runtime: {type(e).__name__}: {e}")
+        print("[LoRA] If you see this, the pipeline may require a restart or an explicit fusion API to apply LoRAs.")
+# ----------------------------------------------------------------
 
 # Preload all models for ZeroGPU tensor packing.
 print("Preloading all models (including Gemma and audio components)...")
@@ -347,7 +407,7 @@ def on_highres_toggle(first_image, last_image, high_res):
     return gr.update(value=w), gr.update(value=h)
 
 
-@spaces.GPU(duration=75)
+@spaces.GPU(duration=80)
 @torch.inference_mode()
 def generate_video(
     first_image,
@@ -360,6 +420,9 @@ def generate_video(
     randomize_seed: bool = True,
     height: int = 1024,
     width: int = 1536,
+    pose_strength: float = 0.0,
+    general_strength: float = 0.0,
+    motion_strength: float = 0.0,
     progress=gr.Progress(track_tqdm=True),
 ):
     try:
@@ -399,6 +462,8 @@ def generate_video(
 
         log_memory("before pipeline call")
 
+        apply_loras_to_pipeline(pose_strength, general_strength, motion_strength)
+
         video, audio = pipeline(
             prompt=prompt,
             seed=current_seed,
@@ -464,9 +529,23 @@ with gr.Blocks(title="LTX-2.3 Heretic Distilled") as demo:
                 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 (set to 0 to disable)")
+                    pose_strength = gr.Slider(
+                        label="Pose Enhancer strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    general_strength = gr.Slider(
+                        label="General Enhancer strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
+                    motion_strength = gr.Slider(
+                        label="Motion Helper strength",
+                        minimum=0.0, maximum=2.0, value=0.0, step=0.01
+                    )
 
         with gr.Column():
-            output_video = gr.Video(label="Generated Video", autoplay=True)
+            output_video = gr.Video(label="Generated Video", autoplay=False)
 
     gr.Examples(
         examples=[
@@ -486,11 +565,15 @@ with gr.Blocks(title="LTX-2.3 Heretic Distilled") as demo:
                 True,
                 1024,
                 1024,
+                0.0,  # pose_strength (example)
+                0.0,  # general_strength (example)
+                0.0,  # motion_strength (example)
             ],
         ],
         inputs=[
             first_image, last_image, input_audio, prompt, duration,
             enhance_prompt, seed, randomize_seed, height, width,
+            pose_strength, general_strength, motion_strength,
         ],
     )
 
@@ -517,6 +600,7 @@ with gr.Blocks(title="LTX-2.3 Heretic Distilled") as demo:
         inputs=[
             first_image, last_image, input_audio, prompt, duration, enhance_prompt,
             seed, randomize_seed, height, width,
+            pose_strength, general_strength, motion_strength,
         ],
         outputs=[output_video, seed],
     )