---
license: mit
tags:
- onnx
- perceptual-hashing
- dinohash
---

# DINOHash — extra checkpoints

Additional DINOHash perceptual-hashing models not present in
[`backslashh/DINOHash`](https://huggingface.co/backslashh/DINOHash).
Each model is provided both as the **raw** training/traced artifact (`raw/`)
and as an exported **ONNX** graph (repo root, dynamic batch axis, opset 17).

| Model | ONNX | Raw | Notes |
|---|---|---|---|
| ViT-Small → ViT-Tiny (DINO distill) | `ViT-Small-ViT-Tiny.onnx` | `raw/ViT-Small-ViT-Tiny.pth` | student backbone (`vit_tiny_patch16_224`), 192-d embedding |
| XCiT-Small → XCiT-Tiny (DINO distill) | `XCiT-Small-XCiT-Tiny.onnx` | `raw/XCiT-Small-XCiT-Tiny.pth` | student backbone (`xcit_tiny_12_p16_224`), 192-d embedding |
| MAE-Lite mae_tiny_400e | `mae_tiny_400e_traced.onnx` | `raw/mae_tiny_400e_traced.pt` | 192-d |
| MAE-Lite mae_tiny_distill_400e | `mae_tiny_distill_400e_traced.onnx` | `raw/mae_tiny_distill_400e_traced.pt` | 192-d |
| MAE-Lite mae_tiny_distill_d2_400e | `mae_tiny_distill_d2_400e_traced.onnx` | `raw/mae_tiny_distill_d2_400e_traced.pt` | 192-d |
| MAE-Lite mocov3_tiny_400e | `mocov3_tiny_400e_traced.onnx` | `raw/mocov3_tiny_400e_traced.pt` | 192-d |

## Notes on the raw files
- **MAE-Lite** raw files are TorchScript (`_traced.pt`), self-contained and loadable directly.
- **ViT / XCiT** raw files are full DINO training checkpoints (`student`/`teacher`/`optimizer`/...).
  The ONNX graphs were built by extracting the `student.backbone.*` weights into the matching
  `timm` architecture (strict-clean load) and exporting; XCiT required `pos_embeder`→`pos_embed`
  rename and qkv split/fuse between class-attention and XCA blocks.

All inputs are `(batch, 3, 224, 224)`.