`Sharp predict` detects the same file twice

[nabagaca]

System Information

• OS: Windows 11
• Python: 3.13

Problem Description

When I run sharp predict, it appears to detect and process the same image twice. Below is a sample of output demonstrating this;

sharp predict -i input_images -o output_splats
| INFO | Processing 2 valid image files.
| INFO | Using device cuda
| INFO | No checkpoint provided. Downloading default model from https://ml-site.cdn-apple.com/models/sharp/sharp_2572gikvuh.pt
| INFO | Using preset ViT dinov2l16_384.
| INFO | Using preset ViT dinov2l16_384.
| INFO | Processing input_images\PXL_20250803_095635868.MP.jpg
| INFO | Running preprocessing.
| INFO | Running inference.
| INFO | Running postprocessing.
| INFO | Saving 3DGS to output_splats
| INFO | Processing input_images\PXL_20250803_095635868.MP.jpg
| INFO | Running preprocessing.
| INFO | Running inference.
| INFO | Running postprocessing.
| INFO | Saving 3DGS to output_splats

I have removed the timestamps from the logs for brevity, but the first processing line is timestamped at 20:11:03,618 and the second at 2025-12-24 20:11:10,267, so 7 seconds between, which leads me to believe this is sharp reprocessing the same image, and not something like the logs/output being duplicated.

I did throw OpenAI Codex at the issue, and it came up with a patch that, as far as I can tell has fixed the issue for me. It seems to solve it by taking the set of files for a given sub-directory.

diff --git a/src/sharp/cli/predict.py b/src/sharp/cli/predict.py
index 8914bb5..bab30a8 100644
--- a/src/sharp/cli/predict.py
+++ b/src/sharp/cli/predict.py
@@ -84,15 +84,26 @@ def predict_cli(
     """Predict Gaussians from input images."""
     logging_utils.configure(logging.DEBUG if verbose else logging.INFO)
 
-    extensions = io.get_supported_image_extensions()
+    extensions = {ext.lower() for ext in io.get_supported_image_extensions()}
 
     image_paths = []
     if input_path.is_file():
-        if input_path.suffix in extensions:
+        if input_path.suffix.lower() in extensions:
             image_paths = [input_path]
     else:
-        for ext in extensions:
-            image_paths.extend(list(input_path.glob(f"**/*{ext}")))
+        seen = set()
+        for candidate_path in input_path.rglob("*"):
+            if not candidate_path.is_file():
+                continue
+            if candidate_path.suffix.lower() not in extensions:
+                continue
+
+            resolved = candidate_path.resolve()
+            if resolved in seen:
+                continue
+
+            seen.add(resolved)
+            image_paths.append(candidate_path)
 
     if len(image_paths) == 0:
         LOGGER.info("No valid images found. Input was %s.", input_path)

A sample of the output after this patch;

sharp predict -i input_images -o output_splats
| INFO | Processing 1 valid image files.
| INFO | Using device cuda
| INFO | No checkpoint provided. Downloading default model from https://ml-site.cdn-apple.com/models/sharp/sharp_2572gikvuh.pt
| INFO | Using preset ViT dinov2l16_384.
| INFO | Using preset ViT dinov2l16_384.
| INFO | Processing input_images\PXL_20250803_095635868.MP.jpg
| INFO | Running preprocessing.
| INFO | Running inference.
| INFO | Running postprocessing.
| INFO | Saving 3DGS to output_splats

As you can see, it correctly detects and processes one image