Run FocalNet's published human-ranking model

[TorstenDittmann]

Ship FocalNet as a selectable Autogravity backend using the published human-ranking ONNX graph, not an importance-only subset.

What this does

• Adds MODEL_BACKEND=focalnet (default remains u2net).
• Loads format-v2 focalnet-human.onnx from the FocalNet 2026-09-14-rc1 release (image, boxes, content → importance, crop_scores).
• Generates FocalNet's candidate crops (5 positions × scales 1.0/0.9/0.8/0.7/0.65), applies the 0.05 importance-retention gate, and returns the selected crop center as gravity.
• Adds a crop rectangle to FocalNet /analyze responses. Optional ?aspect_ratio=16:9 (default 1:1).
• Skips YuNet on this path; faces stay fused in the map.
• Downloads the 19 MiB weights with make model-focalnet and copies them into the Docker image. They are not vendored in git.

make model-focalnet
MODEL_BACKEND=focalnet ./autogravity
# or
docker run --rm -p 8080:8080 -e MODEL_BACKEND=focalnet autogravity

Tests

• Candidate generation, letterbox content, and score_crop match FocalNet Python goldens.
• Retention-gated ranking prefers a high human score among crops that keep enough importance.
• HTTP handler feeds the 3-input graph, returns source=focalnet plus crop, and does not call YuNet.
• Integration: format-v2 dummy ONNX returns a near-center crop on rose.png.
• Local smoke against the published focalnet-human.onnx: 1:1 and 16:9 crops on rose.png.

Unit/vet/fmt passed. Dummy integration and the published human model both ran against ONNX Runtime 1.23.2.

Follow-up

Keep YuNet + U²-Net as the production default until FocalNet is validated in Autogravity traffic. Then flip MODEL_BACKEND in the image.

[appwrite]

Autogravity

Project ID: 6a9e5d640016fa6fb653

Sites (1)

Site	Status	Logs	Preview	QR
Docs autogravity-docs	Ready	View Logs	Preview URL

Tip

HTTPS and SSL certificates are handled automatically for all your Sites

[greptile-apps]

Confidence Score: 4/5

The runtime changes appear sound, but the implementation-coupled FocalNet golden tests violate an explicit repository requirement and must be replaced before merging.

Findings

1.  Golden Tests Mirror Implementation ▶

Fix with agent prompt

### Issue 1
internal/focalnet/candidates_test.go:43-57
This test snapshots every candidate coordinate and its ordering from the Python implementation instead of validating user-visible crop selection. The normalized letterbox assertions in this file and the exact letterbox assertions in `internal/focalnet/letterbox_test.go` repeat the same problem. These tests can fail after a harmless implementation change while still passing when the model selects the wrong subject. This violates the repository directive to test observable behavior rather than mirror implementation details. Replace these snapshots with image-sensitive crop-selection checks against independently annotated subject regions; this repository requirement must be satisfied before merging.

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

---

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

Summary

Adds a selectable FocalNet backend that loads the published human-ranking ONNX graph, generates and ranks crop candidates, and returns the selected crop and normalized gravity point.

• Adds FocalNet preprocessing, inference, map restoration, candidate generation, and retention-gated ranking.
• Extends /analyze with aspect-ratio selection and FocalNet crop metadata.
• Adds model download, checksum verification, optional CI packaging, and mandatory release packaging.
• Keeps U²-Net and YuNet as the default production path.

_{Reviews (5) · Last reviewed commit: "feat: run the published FocalNet human-r..."}