FocalNet 2026-09-14 RC1

First release candidate of the compact FocalNet crop-ranking model.

The same files are published on Hugging Face.

Model

• RepViT-M0.9 importance encoder and lightweight feature pyramid
• 4,757,694 parameters
• 19.45 MiB FP32 ONNX ranking artifact
• 64×64 importance map and scores for up to 128 crop candidates
• Fixed batch size of one, exported with ONNX opset 18

Training

The importance model was trained on 500,000 Open Images samples labeled by an Autogravity teacher using U²-Net and YuNet. The 7,069-parameter composition head was then pretrained on 10,797 CPC images and fine-tuned on GAICD human crop ratings.

Reference results

• GAICD Spearman correlation: 0.76450
• GAICD Pearson correlation: 0.79192
• GAICD top-5 accuracy: 43.70%
• GAICD pairwise accuracy: 86.44%
• Apple Silicon, four threads: 11.69 ms median model time and 20.96 ms median full pipeline time

Assets

Inference:

• focalnet-human.onnx: ranking model
• focalnet-human.onnx.json: ranking preprocessing contract
• focalnet.onnx: importance-only model
• focalnet.onnx.json: importance preprocessing contract

Training checkpoints:

• focalnet-human.pt
• focalnet.pt

Examples from Autogravity:

• comparison-retriever.png

• comparison-faces.png

• MODEL_CARD.md: Hugging Face model card

• focalnet-2026-09-14-rc1-SHA256SUMS: integrity checksums

CPC and GAICD archives do not state explicit image or annotation licenses; these weights do not grant rights to those datasets.