DrayPrescot/bioclip2-onnx

BioCLIP 2 image encoder (ONNX, fp16)

The image encoder of BioCLIP 2 by the Imageomics Institute, converted to ONNX. It is used by Lumenkorn, an open-source photo manager, to name animal species in photos offline.

This is a format conversion only. The weights are those of imageomics/bioclip-2 (ViT-L/14), stored as float16 to halve the download. All credit for the model goes to its authors; please cite their work (below).

Files

File	Size	SHA-256
bioclip2-vision-fp16.onnx	608,910,301 bytes	db55173badd124cdb8ffc48708d4c7db8f1130ad7365baa221bfcc6b66043527
bioclip2-tierarten.bin	8,631,414 bytes	4148db5da336573145985c71561188f73418ea8b6bcdba9aae2591030c02345e

• Input pixel_values: float32, shape [batch, 3, 224, 224]. Preprocess as in OpenCLIP: resize the short side to 224 (bicubic), center crop 224 × 224, scale to 0–1, normalise with mean [0.48145466, 0.4578275, 0.40821073] and std [0.26862954, 0.26130258, 0.27577711].
• Output embedding: float32, shape [batch, 768], L2-normalised.
• The text encoder is not included. Text embeddings (e.g. for a list of taxa) can be computed once with the original model in OpenCLIP and compared by cosine similarity.

bioclip2-tierarten.bin is the species list Lumenkorn uses with this model: about 10,000 animal species (the most observed on iNaturalist in Europe, birds worldwide, all primates, and mammals, birds and reptiles of Kenya, Tanzania and South Africa) with their classes and families, German and English names from Wikidata (CC0), and text embeddings of their taxonomic names computed with the original BioCLIP 2 text encoder (int8 with a scale per entry). The format and the scripts that build it are in tools/bioclip.

Checked against the original PyTorch model in OpenCLIP: cosine similarity of the embeddings ≥ 0.9999 on test photos. export_onnx.py shows how the file was made.

License

MIT, as the original model. See LICENSE (copied from the BioCLIP 2 repository).

Citation

@software{Gu_BioCLIP_2_model,
  author = {Jianyang Gu and Samuel Stevens and Elizabeth G Campolongo and Matthew J Thompson and Net Zhang and Jiaman Wu and Andrei Kopanev and Zheda Mai and Alexander E. White and James Balhoff and Wasila M Dahdul and Daniel Rubenstein and Hilmar Lapp and Tanya Berger-Wolf and Wei-Lun Chao and Yu Su},
  license = {MIT},
  title = {{BioCLIP 2}},
  url = {https://huggingface.co/imageomics/bioclip-2},
  version = {1.0.0},
  doi = {10.57967/hf/5765},
  publisher = {Hugging Face},
  year = {2025}
}

@inproceedings{NEURIPS2025_94da80cb,
  author = {Gu, Jianyang and Stevens, Sam and Campolongo, Elizabeth and Thompson, Matthew and Zhang, Net and Wu, Jiaman and Kopanev, Andrei and Mai, Zheda and White, Alexander and Balhoff, James and Dahdul, Wasila and Rubenstein, Daniel and Lapp, Hilmar and Berger-Wolf, Tanya and Chao, Wei-Lun (Harry) and Su, Yu},
  booktitle = {Advances in Neural Information Processing Systems},
  title = {BioCLIP 2: Emergent Properties from Scaling Hierarchical Contrastive Learning},
  volume = {38},
  pages = {102778--102811},
  year = {2025}
}