Upload 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  = 85
-TEMPORAL_WIN = 60
+MAX_VID_LEN  = 197
+TEMPORAL_WIN = 85
 FPS          = 12
 WEIGHT_DTYPE = torch.bfloat16
 NEG_PROMPT = (
@@ -120,8 +120,8 @@ def load_quadmask_tensor(path: str) -> torch.Tensor:
 
     Quadmask pixel values:
       0   → primary object (to erase)
-      63  → overlap of affected region/ primary object
-      127 → affected region (shadows, reflections, new trajectory)
+      63  → overlap / interaction zone
+      127 → affected region (shadows, reflections …)
       255 → background (keep)
 
     After quantisation the mask is inverted so 255 = "erase", 0 = "keep",
@@ -216,23 +216,41 @@ Use the **VLM-Mask-Reasoner** pipeline included in the repo to generate quadmask
 SAMPLE_DIR = os.path.join(os.path.dirname(__file__), "sample")
 EXAMPLES = [
     [
-        os.path.join(SAMPLE_DIR, "lime",        "input_video.mp4"),
-        os.path.join(SAMPLE_DIR, "lime",        "quadmask_0.mp4"),
+        os.path.join(SAMPLE_DIR, "lime",          "input_video.mp4"),
+        os.path.join(SAMPLE_DIR, "lime",          "quadmask_0.mp4"),
         "A lime falls on the table.",
         30, 1.0, 42,
     ],
     [
-        os.path.join(SAMPLE_DIR, "moving_ball", "input_video.mp4"),
-        os.path.join(SAMPLE_DIR, "moving_ball", "quadmask_0.mp4"),
+        os.path.join(SAMPLE_DIR, "moving_ball",   "input_video.mp4"),
+        os.path.join(SAMPLE_DIR, "moving_ball",   "quadmask_0.mp4"),
         "A ball rolls off the table.",
         30, 1.0, 42,
     ],
     [
-        os.path.join(SAMPLE_DIR, "pillow",      "input_video.mp4"),
-        os.path.join(SAMPLE_DIR, "pillow",      "quadmask_0.mp4"),
+        os.path.join(SAMPLE_DIR, "pillow",        "input_video.mp4"),
+        os.path.join(SAMPLE_DIR, "pillow",        "quadmask_0.mp4"),
         "Two pillows placed on the table.",
         30, 1.0, 42,
     ],
+    [
+        os.path.join(SAMPLE_DIR, "bowling",       "input_video.mp4"),
+        os.path.join(SAMPLE_DIR, "bowling",       "quadmask_0.mp4"),
+        "Bowling pins standing on the grass.",
+        30, 1.0, 42,
+    ],
+    [
+        os.path.join(SAMPLE_DIR, "crush-can",     "input_video.mp4"),
+        os.path.join(SAMPLE_DIR, "crush-can",     "quadmask_0.mp4"),
+        "An empty table.",
+        30, 1.0, 42,
+    ],
+    [
+        os.path.join(SAMPLE_DIR, "toast-shmello", "input_video.mp4"),
+        os.path.join(SAMPLE_DIR, "toast-shmello", "quadmask_0.mp4"),
+        "A plate on the table.",
+        30, 1.0, 42,
+    ],
 ]
 
 with gr.Blocks(title="VOID – Video Object & Interaction Deletion") as demo:
@@ -243,7 +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.
+and VOID will erase the object along with its physical interactions (shadows, deformations, secondary motion).
 
 > Built on **CogVideoX-Fun-V1.5-5B** fine-tuned for interaction-aware video inpainting.
         """