BioCLIP 2

BioCLIP 2 is a CLIP-style vision foundation model for the tree of life: given a photograph of a plant, animal, or fungus, it produces embeddings that support zero-shot and few-shot classification of the organism's species and a wide range of other biological visual tasks. It is the successor to BioCLIP (CVPR 2024) and was developed by the NSF-funded Imageomics Institute, led by The Ohio State University. The model and its accompanying dataset were released in May 2025 and the work was accepted as a spotlight at NeurIPS 2025.

Identifying organisms from images is a foundational problem across ecology, conservation, and systematics, but the space of life is vast and long-tailed: most species are rare, visually similar to close relatives, and poorly represented in conventional image datasets. BioCLIP 2 addresses this by scaling contrastive vision-language training to biodiversity data. The authors first curated TreeOfLife-200M, comprising 214 million images of living organisms spanning 952,000 taxa, then trained BioCLIP 2 with a hierarchical contrastive objective that aligns each image with its full Linnaean taxonomic label rather than a flat class index.

The paper's central finding is that this narrow training objective, distinguishing species, produces emergent behavior. Despite never being trained on ecological or morphological labels, BioCLIP 2 achieves strong accuracy on downstream tasks such as habitat classification and trait prediction, and its embedding space organizes itself along biologically meaningful axes.

#Key Features

• Tree-of-life-scale training: BioCLIP 2 is trained on TreeOfLife-200M, 214 million organism images across 952K taxa, the largest and most taxonomically diverse biological image dataset assembled to date.
• Hierarchical contrastive learning: Rather than treating species as independent labels, the model contrasts images against textual taxonomic hierarchies, encoding the nested structure of biological classification.
• Zero-shot and few-shot classification: The learned image-text embedding space enables classification of species never seen during fine-tuning simply by supplying candidate taxon names, with no task-specific retraining.
• Emergent embedding structure: At the inter-species level, embeddings align with functional and ecological meaning (e.g., beak sizes, habitats); at the intra-species level, variations such as life stage and sex are preserved in subspaces orthogonal to species distinctions.
• Open release: Model weights and code are released under the MIT license and the TreeOfLife-200M dataset under CC0, with model and dataset cards on HuggingFace.

#Technical Details

BioCLIP 2 uses a ViT-L/14 image encoder paired with a masked self-attention Transformer text encoder, fine-tuned from an OpenCLIP checkpoint pretrained on LAION-2B. Training scales up the original BioCLIP recipe with a hierarchical contrastive objective over taxonomic labels, dual visual projectors, and an experience-replay mechanism that mixes in roughly 26 million LAION-2B samples to retain general visual competence while specializing on organisms. In the authors' evaluations, BioCLIP 2 improves species-classification accuracy by 18.1% over the original BioCLIP, reaching 55.6% mean accuracy across zero-shot species benchmarks and 57.5% mean performance on non-species biological visual tasks such as habitat and trait prediction. The paper accompanies these results with formal analysis of why hierarchical supervision and contrastive objectives encourage the observed emergent structure, and shows that the effect strengthens with larger training data.

#Applications

BioCLIP 2 is aimed at ecologists, conservation biologists, systematists, and biodiversity-informatics teams who need to identify and characterize organisms at scale. Its zero-shot and few-shot capabilities suit camera-trap surveys, museum specimen digitization, citizen-science platforms such as iNaturalist, and rapid biodiversity assessment where labeled training data for the target taxa is scarce. Because the embeddings capture ecological and morphological structure beyond species identity, they also support downstream tasks like trait measurement, habitat inference, and organizing large unlabeled image collections.

#Impact

BioCLIP 2 extends the BioCLIP line of biodiversity foundation models and demonstrates that scaling hierarchical contrastive learning yields a biologically meaningful embedding space with capabilities beyond its training objective. Its open release, MIT-licensed weights and code alongside the CC0 TreeOfLife-200M dataset, lowers the barrier for the ecology and computer-vision communities to build on organism-scale representations, and the accompanying dataset is itself a significant community resource. As the model matures, the Imageomics Institute has continued the series with larger variants, positioning BioCLIP 2 as a widely adopted backbone for AI applied to the natural world.