DINOHash
Canonical model zoo for DINOHash — robust perceptual image hashing (ONNX).
All inputs are (N, 3, 224, 224), ImageNet-normalized
(mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225]).
Recommended — hash bits baked in
The ONNX graph includes the PCA head; the output is per-bit logits (binarize with >= 0 → 1).
| Model | Bits | File |
|---|
| DINOv2 ViT-B/14 (flagship) | 512 | dinov2_vitb14_reg_512bit_dynamic.onnx |
| DINOv2 ViT-S/14 | 96 | dinov2_vits14_reg_96bit_dynamic.onnx |
Other backbones — raw embeddings (apply your own PCA / sign-binarization)
- DINOv3:
dinov3-vits16, dinov3-vits16plus, dinov3_convnext-tiny, dinov3_convnext-small
- Distilled students (DisCo):
ResNet-{50,50*2,101,152} → {ResNet-18, ResNet-34, MobileNet-v3, EfficientNet-B0, EfficientNet-B1}, plus ViT-Small→ViT-Tiny, XCiT-Small→XCiT-Tiny
- SSL baselines: DINO / MoCo / SwAV (MobileNetV2, R18, R34, ViT-T) and MAE-Lite (
mae_tiny_*, mocov3_tiny)
raw/
Raw training / TorchScript checkpoints for the ViT-Tiny, XCiT-Tiny and MAE-Lite models.
