Use non-overlapping full blocks with boundary-only completion

README.md

@@ -51,7 +51,7 @@ do not represent eight entirely held-out generators.
 
 ## Download and run
 
-Use **Python 3.11**. Replace `MuskAI/LTF-v3.3` with this repository's ID.
+Use **Python 3.11**. The commands below use this model repository.
 
 ```bash
 python -m pip install -U huggingface_hub
@@ -81,8 +81,8 @@ print(result["prediction"], result["score_fake"])
 print(result["probabilities"])
 ```
 
-The image is decoded as RGB. All original edge-aligned 224×224 local tiles plus
-a global view use the original CLIP preprocessing. Input files are not
+The image is decoded as RGB. Non-overlapping full 224×224 local blocks, with right/bottom boundary completion,
+plus a global view use CLIP preprocessing. Input files are not
 recompressed. Test tiling depends on image size. Labels are **0=real, 1=fake**.
 The API returns L1/L2/fusion scores, the selected label and the view count.
 Loading the model does not download external weights after this repository
@@ -163,3 +163,12 @@ Please cite the LTF paper:
   archivePrefix={arXiv}
 }
 ```
+
+## Tiling policy
+
+Full 224×224 blocks are taken without overlap. If pixels remain at the right
+or bottom, a final block is aligned to that boundary; only boundary blocks
+may overlap. For width 500, horizontal starts are 0, 224, 276. The global view
+is separate. This corrects earlier uniformly overlapping tiling; weights are
+unchanged, but predictions can change. Historical benchmark results have not
+been recomputed with this preprocessing.

demo/app.py

@@ -495,7 +495,7 @@ def build_demo():
                         elem_classes="attention-image",
                     )
                     gr.Markdown(
-                        "Tiles are numbered row by row. Edge alignment can produce overlap. Some labels are omitted on dense grids; complete coordinates are listed below."
+                        "Tiles are numbered row by row. Full 224-pixel blocks do not overlap; only the final right/bottom boundary blocks may overlap. Some labels are omitted on dense grids; complete coordinates are listed below."
                     )
                 with gr.Tab("Tile · Patch Details", id="patch"):
                     gr.Markdown(

demo/attention.py

@@ -64,7 +64,7 @@ def capture_prediction(detector, path, branch, threshold):
 
 
 def tile_boxes(width, height, tile_size=224):
-    # Use the exact same edge-aligned coordinates and row-major order as inference.
+    # Match inference: non-overlapping full blocks, then right/bottom boundary blocks.
     from utils.tile_dataset_srm import get_start_positions
 
     return [
@@ -256,6 +256,7 @@ def parameter_rows(detector, branch, threshold, views):
         ],
         ["Inference", "CLIP chunk size", str(detector.model.eval_chunk_size)],
         ["Inference", "Views: global + local", f"1 + {views - 1} = {views}"],
+        ["Inference", "Tiling policy", "Non-overlapping full blocks; overlap only at right/bottom boundaries"],
         ["Architecture", "Backbone", "Frozen CLIP ViT-L/14"],
         [
             "Architecture",

ltf_final/detector.py

@@ -1,4 +1,4 @@
-"""Offline model loading and the original experiment's image preprocessing."""
+"""Offline model loading and non-overlapping tiling with boundary completion."""
 from __future__ import annotations
 
 import contextlib
@@ -78,7 +78,7 @@ def create_model(config, weights=None, device='cpu'):
 
 
 def prepare_image(path, tile_size=224):
-    """Same RGB, tiling, Resize/CenterCrop and CLIP normalization as original eval."""
+    """Original RGB and CLIP normalization; non-overlapping blocks with boundary completion."""
     path = Path(path).expanduser().resolve()
     with Image.open(path) as source:
         image = source.convert('RGB')
@@ -122,4 +122,5 @@ class LTFDetector:
         label = int(score >= threshold)
         return dict(path=str(path), prediction='fake' if label else 'real', label=label,
                     branch=branch, threshold=float(threshold), score_fake=score,
+                    tiling_policy="nonoverlap_then_boundary",
                     views=int(images.shape[0]), probabilities=probabilities)

release_manifest.json

@@ -11,7 +11,8 @@
     "Tensor-only checkpoint loading",
     "Portable evaluation configuration",
     "Preserve Hub snapshot root when source files are symlinks",
-    "English source comments and docstrings; executable AST unchanged"
+    "English source comments and docstrings; executable AST unchanged",
+    "Non-overlapping full 224-pixel blocks with right/bottom boundary completion; historical benchmark numbers not recomputed"
   ],
   "files": {
     "ltf_final/__init__.py": {
@@ -19,8 +20,8 @@
       "bytes": 175
     },
     "ltf_final/detector.py": {
-      "sha256": "998a82c15b253b52905c02201d79b7d92fff30fa95a95c97d48772206d44fb27",
-      "bytes": 6074
+      "sha256": "fee36f9976efea2de60f0d106f9a3693ebc394cfa5dd4277470bd618f0612307",
+      "bytes": 6147
     },
     "model/LTF.py": {
       "sha256": "60d12ab01bf5d748ab0d0d453d99e18512d1a5043c742a2030108d9f15cb7f92",
@@ -63,8 +64,8 @@
       "bytes": 0
     },
     "utils/tile_dataset_srm.py": {
-      "sha256": "49106f01b4fda4d10e7db8364967988a68a00c39daa2dde8fe89f52506531642",
-      "bytes": 14623
+      "sha256": "a79a5edc0804aa0331805ba27276d09fa4928b3f2938a08eb3fa398682824c84",
+      "bytes": 14576
     },
     "configs/eval.yaml": {
       "sha256": "4ce443731e70a4ac4eb71fb56103f33373e179fefe766c3e6a02635efdb94efc",
@@ -78,5 +79,6 @@
       "sha256": "b8cca3fd41ae0c99ba7e8951adf17d267cdb84cd88be6f7c2e0eca1737a03836",
       "bytes": 932768134
     }
-  }
+  },
+  "tiling_policy": "nonoverlap_then_boundary"
 }

utils/tile_dataset_srm.py

@@ -52,15 +52,13 @@ def pad_crop_image(image, target_width, target_height):
     return image
 
 def get_start_positions(image_length, block_size, bias=0):
+    """Non-overlapping full blocks, plus one boundary block when needed."""
     if image_length <= block_size:
-        return [0+bias]
-    num_blocks = image_length // block_size + (image_length % block_size > 0)
-    if num_blocks == 1:
-        return [0+bias]
-    total_stride = (image_length - block_size) / (num_blocks - 1)
-    positions = [int(round(i * total_stride) + bias) for i in range(num_blocks - 1)]
-    positions.append(image_length - block_size + bias)
-    return positions
+        return [bias]
+    positions = list(range(0, image_length - block_size + 1, block_size))
+    if positions[-1] != image_length - block_size:
+        positions.append(image_length - block_size)
+    return [position + bias for position in positions]
 
 def crop_image_to_tiles(image, block_size):
     """