Publish FocalNet checkpoints on Hugging Face

[TorstenDittmann]

Pretrained FocalNet weights are published to the existing Hub repo appwrite/focalnet and attached to the 2026-09-14-rc1 GitHub release.

What changed

• Added a versioned model card at docs/huggingface-model-card.md and scripts/publish-huggingface.py, which refuses to upload unless SHA-256 matches the documented reference hashes, hashes sit next to the matching filenames, example images exist, and ONNX sidecars match embedded metadata.
• README, result notes, and the crop lab fetch focalnet-human.onnx from Hugging Face by default (bun run fetch-model) and verify the published hash. An empty FOCALNET_ONNX_SHA256 still verifies.
• Hub examples are stacked source/center/FocalNet figures rendered from Autogravity photographs. The renderer takes model, photo, output, and font paths and falls back to DejaVu/Liberation when Inter is missing.
• The GitHub release keeps the model files: importance and ranking ONNX, JSON sidecars, PyTorch checkpoints, and checksums. Example images and the model card stay on Hugging Face.
• Download examples use curl / urllib so a runtime-only pip install . environment can fetch weights without Hugging Face CLI packages.
• Weights stay gitignored. CPC/GAICD dataset-rights caveats stay in the card and README.

  

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge despite a non-blocking defect in the exported fetch helper’s custom-destination dummy fallback.

Findings

1.  Dummy Ignores Custom Destination ▶

Fix with agent prompt

### Issue 1
web/scripts/fetch-model.ts:109-117
When a caller enables dummy mode with a non-default `destPath`, the exporter still writes the model to the hardcoded `web/public/focalnet-human.onnx`. `fetchModel` then hashes the requested path, which was never created, so the call rejects after claiming it wrote the dummy. Pass the requested destination to the exporter or create the dummy directly at `destPath`.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

Summary

This PR publishes verified FocalNet checkpoints and supporting model-card assets, adds guarded Hugging Face staging and upload tooling, makes the browser model fetcher use the published ranking checkpoint by default, and adds rendering and publishing tests.

• Checkpoint hashes and ONNX sidecar metadata are validated before publication.
• Model-card image and filename/hash associations are checked during staging.
• Browser downloads now verify the default checkpoint hash, including when the hash environment variable is empty.
• Runtime-only download examples no longer require Hugging Face CLI dependencies.
• The dummy-model fallback mishandles the newly configurable destination path.

_{Reviews (4) · Last reviewed commit: "Verify ONNX sidecars after all Hub files..."}