Update app.py

app.py

@@ -291,7 +291,7 @@ pipeline = LTX23DistilledA2VPipeline(
             LTXV_LORA_COMFY_RENAMING_MAP
         )
     ],
-    quantization=QuantizationPolicy.fp8_cast(),
+    quantization=None,
 )
 
 # Preload all models for ZeroGPU tensor packing.
@@ -303,12 +303,26 @@ _original_forward = _transformer.forward
 def _lora_scaled_forward(*args, **kwargs):
     out = _original_forward(*args, **kwargs)
 
-    # Apply runtime scaling to LoRA-influenced output
-    # (LTX merges LoRA into attention residuals, so we scale output delta)
+    # Only scale deviation from baseline (approximation)
+    scale = LORA_RUNTIME_SCALE
+
+    if scale == 1.0:
+        return out
+    if scale == 0.0:
+        # crude fallback: suppress output magnitude slightly
+        if torch.is_tensor(out):
+            return out * 0.5
+        return out
+
+    if torch.is_tensor(out):
+        return out * scale
+
     if isinstance(out, tuple):
-        return tuple(o * LORA_RUNTIME_SCALE if torch.is_tensor(o) else o for o in out)
-    elif torch.is_tensor(out):
-        return out * LORA_RUNTIME_SCALE
+        return tuple(
+            o * scale if torch.is_tensor(o) else o
+            for o in out
+        )
+
     return out
 
 _transformer.forward = _lora_scaled_forward
@@ -519,6 +533,7 @@ with gr.Blocks(title="LTX-2.3 Heretic Distilled") as demo:
                 "crowd together closely, forming a symmetrical cluster while staring "
                 "directly into the lens.",
                 3.0,
+                1.0,
                 False,
                 42,
                 True,
@@ -527,7 +542,7 @@ with gr.Blocks(title="LTX-2.3 Heretic Distilled") as demo:
             ],
         ],
         inputs=[
-            first_image, last_image, input_audio, prompt, duration,
+            first_image, last_image, input_audio, prompt, duration, lora_strength,
             enhance_prompt, seed, randomize_seed, height, width,
         ],
     )