Update app.py

app.py

@@ -39,8 +39,8 @@ VOID_MODEL_ID  = os.environ.get("VOID_MODEL_ID", "your-hf-username/VOID")
 VOID_CKPT_FILE = "void_pass1.safetensors"
 
 SAMPLE_SIZE  = (384, 672)   # H × W
-MAX_VID_LEN  = 197
-TEMPORAL_WIN = 85
+MAX_VID_LEN  = 72
+TEMPORAL_WIN = 60
 FPS          = 12
 WEIGHT_DTYPE = torch.bfloat16
 NEG_PROMPT = (
@@ -206,8 +206,8 @@ The quadmask is a **grayscale video** where each pixel value encodes what role t
 | Pixel value | Meaning |
 |-------------|---------|
 | **0** (black) | Primary object to remove |
-| **63** (dark grey) | Overlap / interaction zone |
-| **127** (mid grey) | Affected region — shadows, reflections, secondary effects |
+| **63** (dark grey) | Overlap of primary object / affected zone |
+| **127** (mid grey) | Affected region — shadows, reflections, new and old trajectories |
 | **255** (white) | Background — keep as-is |
 
 Use the **VLM-Mask-Reasoner** pipeline included in the repo to generate quadmasks automatically.
@@ -261,7 +261,7 @@ with gr.Blocks(title="VOID – Video Object & Interaction Deletion") as demo:
 [🌐 Project Page](https://void-model.github.io/) | [💻 GitHub](https://github.com/void-model/VOID)
 
 Upload a video and its **quadmask**, enter a prompt describing the scene *after* removal,
-and VOID will erase the object along with its physical interactions (shadows, deformations, secondary motion).
+and VOID will erase the object along with its physical interactions.
 
 > Built on **CogVideoX-Fun-V1.5-5B** fine-tuned for interaction-aware video inpainting.
         """