---
license: other
license_name: per-model
license_link: LICENSE.md
library_name: onnx
pipeline_tag: image-feature-extraction
tags:
  - onnx
  - onnxruntime-web
  - webgpu
  - timm
  - explainability
---

# Demo models

Vision models converted to ONNX for the in-browser demos on [karay.me](https://karay.me):
the [LaFAM demo](https://karay.me/examples/lafam/) and the
[region similarity demo](https://karay.me/examples/similarity/). The demos run them on the visitor's
device with [ONNX Runtime Web](https://onnxruntime.ai/) (WebGPU, or WebAssembly on the CPU); no image
leaves the browser.

Each file is the largest model of its family that stays under 100 MB, except where a specific model is the
point: CLIP RN50x4 (the model of the typographic attacks of Goh et al., 2021) and the pathology models
Hibou-B and QuiltNet-B-32.

## Interface

Every model takes one input and returns the last feature map before pooling:

| | Shape | Meaning |
|---|---|---|
| input `image` | (1, 3, S, S) | RGB in [0, 1]; S = 224 (288 for CLIP RN50x4); the model's own normalization is part of the graph |
| output `features` | (1, C, H, W) | the last feature map; for ViTs, the patch tokens after the final norm, as a grid |
| output `cell_logits` | (1, K, H, W) | classifiers only: the classification head applied to every cell (K = 1000 for ImageNet, 5 tissue classes, 2 for the tumour detector) |
| output `logits` | (1, K) | only CLIP RN50x4 and QuiltNet-B-32: the model's own zero-shot logits (attention pooling or class token), which are not the mean of the cells |

For the classifiers that pool their features (ResNet, ConvNeXt, Swin, MaxViT), the mean of
`cell_logits` over the cells equals the model's logits, so the per-cell readout is exact
([Karjauv, 2026](https://arxiv.org/abs/2606.14555)). For the two CLIP models with a zero-shot head, `cell_logits`
classifies each cell on its own: the cell's value embedding through the attention pooling's output
projection (RN50x4, as in MaskCLIP, [Zhou et al., 2022](https://arxiv.org/abs/2112.01071)), or the patch
token through the final projection (QuiltNet). Weights are stored as float16 and cast to float32
in the graph; the files match the PyTorch models to a correlation of at least 0.9999 on the demo images.

## Files

| File | Model | Source checkpoint (timm) | Grid | License |
|---|---|---|---|---|
| `resnet50.onnx` | ResNet-50, ImageNet classifier | [`resnet50.tv2_in1k`](https://huggingface.co/timm/resnet50.tv2_in1k) | 7×7 | BSD-3-Clause |
| `convnext_small.onnx` | ConvNeXt-S, ImageNet classifier | [`convnext_small.fb_in1k`](https://huggingface.co/timm/convnext_small.fb_in1k) | 7×7 | Apache-2.0 |
| `swin_small.onnx` | Swin-S, ImageNet classifier | [`swin_small_patch4_window7_224.ms_in1k`](https://huggingface.co/timm/swin_small_patch4_window7_224.ms_in1k) | 7×7 | MIT |
| `maxvit_tiny.onnx` | MaxViT-T, ImageNet classifier | [`maxvit_tiny_tf_224.in1k`](https://huggingface.co/timm/maxvit_tiny_tf_224.in1k) | 7×7 | Apache-2.0 |
| `dino_resnet50.onnx` | DINO ResNet-50, self-supervised | [DINO](https://github.com/facebookresearch/dino) `dino_resnet50_pretrain.pth` (torchvision) | 7×7 | Apache-2.0 |
| `dino_vit_small.onnx` | DINO ViT-S/16, self-supervised | [`vit_small_patch16_224.dino`](https://huggingface.co/timm/vit_small_patch16_224.dino) | 14×14 | Apache-2.0 |
| `dinov2_vit_small.onnx` | DINOv2 ViT-S/14, self-supervised | [`vit_small_patch14_dinov2.lvd142m`](https://huggingface.co/timm/vit_small_patch14_dinov2.lvd142m) | 16×16 | Apache-2.0 |
| `dinov2_reg_vit_small.onnx` | DINOv2 ViT-S/14 with registers, self-supervised | [`vit_small_patch14_reg4_dinov2.lvd142m`](https://huggingface.co/timm/vit_small_patch14_reg4_dinov2.lvd142m) | 16×16 | Apache-2.0 |
| `dinov3_vit_small_plus.onnx` | DINOv3 ViT-S+/16, self-supervised | [`vit_small_plus_patch16_dinov3.lvd1689m`](https://huggingface.co/timm/vit_small_plus_patch16_dinov3.lvd1689m) | 14×14 | [DINOv3 License](DINOV3_LICENSE.md) |
| `dinov3_convnext_small.onnx` | DINOv3 ConvNeXt-S, self-supervised | [`convnext_small.dinov3_lvd1689m`](https://huggingface.co/timm/convnext_small.dinov3_lvd1689m) | 7×7 | [DINOv3 License](DINOV3_LICENSE.md) |
| `fcmae_convnextv2_tiny.onnx` | ConvNeXt V2-T, masked autoencoder (FCMAE) | [`convnextv2_tiny.fcmae`](https://huggingface.co/timm/convnextv2_tiny.fcmae) | 7×7 | CC-BY-NC-4.0 (non-commercial use only) |
| `clip_resnet101.onnx` | CLIP ResNet-101 (OpenAI), image encoder trunk | [`resnet101_clip_gap.openai`](https://huggingface.co/timm/resnet101_clip_gap.openai) | 7×7 | MIT |
| `tinyclip_vit_medium.onnx` | TinyCLIP ViT-M/16, image encoder | [`vit_medium_patch16_clip_224.tinyclip_yfcc15m`](https://huggingface.co/timm/vit_medium_patch16_clip_224.tinyclip_yfcc15m) | 14×14 | MIT |
| `clip_rn50x4.onnx` | CLIP RN50x4 (OpenAI) with its zero-shot ImageNet classifier (80 prompts) | open_clip `RN50x4`, `openai` | 9×9 | MIT |
| `hibou_b.onnx` | Hibou-B (pathology, DINOv2 ViT-B/14 with registers) + a linear tissue head (tumour, stroma, lymphocytes, necrosis, fat) | [`histai/hibou-b`](https://huggingface.co/histai/hibou-b) via `hibou_port.py`; head `hibou_b_tissue_head.pt` | 16×16 | Apache-2.0 |
| `quiltnet_b32.onnx` | QuiltNet-B-32 (pathology CLIP) with a zero-shot head over the same tissue classes | open_clip [`wisdomik/QuiltNet-B-32`](https://huggingface.co/wisdomik/QuiltNet-B-32) | 7×7 | MIT |
| `tumour_detector.onnx` | breast tumour detector, ResNet-34 (other, tumour) | [`kaczmarj/breast-tumor-resnet34.tcga-brca`](https://huggingface.co/kaczmarj/breast-tumor-resnet34.tcga-brca) (torchvision) | 7×7 | CC-BY-4.0 |
| `kaiko_vit_small.onnx` | Kaiko ViT-S/16 (pathology, self-supervised) | torch.hub [`kaiko-ai/towards_large_pathology_fms`](https://github.com/kaiko-ai/towards_large_pathology_fms) `vits16` | 14×14 | Kaiko Non-Commercial |
| `ctranspath.onnx` | CTransPath (pathology, Swin-T with a convolutional stem) | [`1aurent/swin_tiny_patch4_window7_224.CTransPath`](https://huggingface.co/1aurent/swin_tiny_patch4_window7_224.CTransPath) | 7×7 | GPL-3.0 |

The pathology models expect H&E tiles at about 0.5 µm per pixel (224 pixels = 112 µm); the tumour detector
was trained on 0.25 µm per pixel tiles of 350 pixels resized to 224 (about 0.39 µm per pixel). The tissue
head of Hibou-B is logistic regression on the mean patch token, trained with one label per 112 µm tile
(the majority tissue of the pathologist's annotation) on 6,692 tiles of 132 BCSS cases, none lobular
(`hibou_head.py` in the author's project); as it is linear on the mean, the mean of its `cell_logits` is
its prediction.

Each file is a derivative of its source checkpoint and is distributed under that checkpoint's license
(see [LICENSE.md](LICENSE.md)). The DINOv3 models are distributed under the DINOv3 License, a copy of
which is included; ConvNeXt V2 (FCMAE) and Kaiko ViT-S/16 are for non-commercial use only; CTransPath is
GPL-3.0, with its conversion source in this repository.

## Reproducing

[`export_models.py`](export_models.py) exports every file from its source checkpoint and checks it
against PyTorch on a folder of .jpg images (any will do):

    SAMPLES=path/to/images uv run --no-project --with torch --with torchvision --with timm --with onnx --with onnxruntime \
        --with pillow --with open_clip_torch --with safetensors --with huggingface_hub python export_models.py NAME