Update code/comments from local workspace

app.py

@@ -6,7 +6,6 @@ from PIL import Image
 import torch
 import spaces
 
-# Disable matplotlib visualizations inside the backend call (Spaces are headless)
 import utils.visualize as vis
 vis.visualize_segmentation = lambda *args, **kwargs: None  # type: ignore
 
@@ -14,14 +13,13 @@ from models.model_bank_knn import PatchKNNDetector
 from backbones import get_backbone
 from segmenters import SAM3Segmenter
 
-# ZeroGPU: avoid initializing CUDA at import time. Keep everything on CPU until the
-# GPU-decorated inference runs and a slice is attached.
+
 DEFAULT_DEVICE = "cpu"
 
 
 @functools.lru_cache(maxsize=1)
 def load_backbone(name: str = "dinov3_small"):
-    # Keep on CPU; move to GPU inside infer when available
+    # Keep on CPU will move to gpu if available
     return get_backbone(name).to(DEFAULT_DEVICE).eval()
 
 
@@ -57,7 +55,7 @@ def _make_overlay(rgb_image: np.ndarray, anomaly_map: np.ndarray) -> Image.Image
     return Image.fromarray(cv2.cvtColor(overlay_bgr, cv2.COLOR_BGR2RGB))
 
 
-@spaces.GPU  # When running on ZeroGPU, this grants a short-lived GPU slice for the call.
+@spaces.GPU
 def infer(ref_files, test_file, use_sam3, sam_prompt):
     if not ref_files:
         raise gr.Error("Upload at least one reference image.")
@@ -103,7 +101,7 @@ def build_demo():
     with gr.Blocks(title="Patch KNN Anomaly Detection") as demo:
         gr.Markdown(
             "# Patch KNN Anomaly Detection\n"
-            "Upload reference (normal) images, one test image, and optionally run SAM3 to focus on a specific object."
+            "Upload reference (good) images, one test image, and optionally run SAM3 to segment the specific foreground object."
         )
 
         with gr.Row():

segmenters/sam3.py

@@ -3,8 +3,8 @@ from __future__ import annotations
 import numpy as np
 import torch
 from PIL import Image
-import os
-from transformers import Sam3Processor, Sam3Model
+import os
+from transformers import Sam3Processor, Sam3Model
 
 from segmenters import BaseSegmenter
 
@@ -41,20 +41,19 @@ class SAM3Segmenter(BaseSegmenter):
         self.score_threshold = score_threshold
         self.mask_threshold = mask_threshold
 
-        # Loading model model + defining processor
-        token = os.getenv("HF_TOKEN")
-        # facebook/sam3 is a gated model; token required on Spaces without pre-approval
-        self.model = Sam3Model.from_pretrained(
-            model_name,
-            token=token,
-            trust_remote_code=True,
-        ).to(self.device)
-        self.model.eval()
-        self.processor = Sam3Processor.from_pretrained(
-            model_name,
-            token=token,
-            trust_remote_code=True,
-        )
+        # Loading model + defining processor
+        token = os.getenv("HF_TOKEN")
+        self.model = Sam3Model.from_pretrained(
+            model_name,
+            token=token,
+            trust_remote_code=True,
+        ).to(self.device)
+        self.model.eval()
+        self.processor = Sam3Processor.from_pretrained(
+            model_name,
+            token=token,
+            trust_remote_code=True,
+        )
 
     def get_object_mask(self, image: np.ndarray) -> np.ndarray:
         """