aloxaf/dinov3-small-onnx

DINOv3 ViT-S/16 — CPU FP32 ONNX

ONNX export of Meta DINOv3 ViT-S/16 pretrained on LVD-1689M. This is an unquantized export, not a fine-tuned model.

• Input: pixel_values, float32 [batch, 3, 224, 224]; dynamic batch, fixed spatial size.
• Output: patch_tokens, float32 [batch, 196, 384]; normalized patch tokens, excluding CLS and four register tokens.
• ONNX opset 18. Validated with ONNX Runtime 1.28 CPUExecutionProvider at batches 1, 8 and 3.
• Intended preprocessing: RGB, resize short edge to 256 with bicubic interpolation, center crop 224, ImageNet mean/std normalization. This preserves the tested storyboard pipeline's view; it differs from the upstream default full-image 256 resize.
• For the tested spatial representation, arrange patch tokens as 14×14, apply adaptive average pooling to 4×4, then L2-normalize each of the 16 cells. Pooling is not included in the ONNX graph.

Source and license

Original model: https://huggingface.co/facebook/dinov3-vits16-pretrain-lvd1689m

Official implementation: https://github.com/facebookresearch/dinov3

Source revision, original checkpoint checksum, export dependency versions and numerical validation are recorded in source.json. The original checkpoint was obtained from the public ciqiangxu/DINOv3 ModelScope redistribution at the recorded revision.

The original DINOv3 License applies; see LICENSE.md. Meta owns the underlying model. This repository contains no training samples, storyboard images, video frames, or credentials.