Model reference · open weights

vit_reg4_so150m_p14_ls_dino-bio

Available as managed deployment Embeddings birder-project Image embed 1 variants 6k dl/mo

vit_reg4_so150m_p14_ls_dino-bio is an open-weight embedding model from birder-project. AxForge deploys and operates it for you on dedicated EU-owned hardware — with the licence handled where one is required.

Available as managed deployment — configured and operated for you on dedicated EU hardware, quoted per deployment.

What it is

Released bybirder-project
TypeEmbedding models
TaskImage embed
Runs withbirder
Based onbirder-project/dino_v2_vit_reg4_so150m_p14_ls_bio
Released2026-05-09
Popularity6k downloads / month
LicenceOpen weights

About

What vit_reg4_so150m_p14_ls_dino-bio is

vit_reg4_so150m_p14_ls_dino-v2-bio is a Bio-DINO image encoder for natural photographs of living organisms. It uses a SoViT-150M/14 Vision Transformer with 4 register tokens and 133.6M backbone parameters, trained with a DINOv2-style self-supervised objective on approximately 31 million curated images spanning plants, fungi, insects, fish, corals, birds, mammals and other biodiversity subjects.

The model is released as a single repository with three checkpoints: 224 x 224, 252 x 252 and 336 x 336. The 252px checkpoint is the recommended default, the 224px checkpoint is a faster and lower-cost variant, and the 336px checkpoint is a short high-resolution adaptation for users who want to trade latency for more input detail. All checkpoints are intended primarily as frozen image encoders for embeddings, retrieval, clustering, probing and transfer learning, rather than as ready-made species classifiers.

This repository releases the Birder backbone checkpoints for direct use as image encoders. The full DINO training weights, including the DINO head and training state needed for continued self-supervised work, are released separately in the full DINO training weights companion repository.

Read the full model card

Model Details

  • Model Type: Image encoder and detection backbone

  • Model Stats:

    • Params (M): 133.6
    • Input image size: 252 x 252
    • Additional released sizes: 224 x 224, 336 x 336
  • Dataset: Trained on a diverse dataset of approximately 31M images, including:

    • TreeOfLife-10M-EOL-NaturalImages
    • iNaturalist 2021
    • BIOSCAN-5M (pretrain split)
    • TreeOfLife-200M (subset)
    • IP102 v1.1
    • iWildCam 2022 (subset)
    • The Birder dataset (private dataset)
  • Papers:

    • An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale:
    • Vision Transformers Need Registers:
    • Getting ViT in Shape: Scaling Laws for Compute-Optimal Model Design:
    • DINOv2: Learning Robust Visual Features without Supervision:

Released Checkpoints / Model Variants

All three backbone checkpoints are released in this model repository.

CheckpointResolutionTraining stageRecommended use
vit_reg4_so150m_p14_ls_dino-v2-bio-224px224 x 224Initial DINOv2 training resolutionFastest Bio-DINO checkpoint, useful when throughput or memory is the main constraint
vit_reg4_so150m_p14_ls_dino-v2-bio-252px252 x 252Continued training from 224pxRecommended default checkpoint for most embedding and transfer workflows
vit_reg4_so150m_p14_ls_dino-v2-bio-336px336 x 336Short high-resolution adaptation from 252pxHigher-input-detail checkpoint, useful when latency and memory are less constrained
Full DINO training weightsvaries-Continued self-supervised training and research

Intended Use And Scope

This model is intended for natural photographs of living organisms and biodiversity subjects: organismal field photography, camera-trap imagery, collection-style photographs, underwater photography and similar real-world images of living systems. It is best used as a general-purpose image encoder for representation-learning workflows, especially when the downstream task benefits from visual similarity, fine-grained biological structure or transfer from broad biodiversity imagery. Image retrieval and visual similarity search are primary use cases for Bio-DINO. The model was evaluated heavily as a frozen embedding model and is particularly well suited to finding visually related biological images across large collections.

Suitable Uses

Suitable uses include:

  • Image embedding extraction for natural photographs of living organisms.
  • Nearest-neighbor retrieval, visual similarity search and reference-image matching.
  • Clustering, dataset exploration, and biodiversity image organization.
  • Dataset deduplication, near-duplicate review, and collection triage.
  • Frozen-embedding evaluation with linear probes, k-NN, SimpleShot, SVM, MLP probes, or retrieval metrics.
  • Lightweight supervised transfer for taxonomic, ecological, trait, or visual-recognition tasks.
  • Fine-tuning or initialization for downstream biodiversity image models.
  • Continued self-supervised training and representation-learning research.

Out-of-Scope Domains

The model is not intended for microscopy, pathology, medical imaging, X-ray/radiology, satellite imagery, maps, diagrams, charts, drawings, document understanding or other non-photographic visual domains. It was trained for natural photography, not controlled lab imagery, scientific instrument imagery or clinical imagery.

Training Data

The model was trained on approximately 31 million curated images focused on natural photographs of living organisms and biodiversity subjects. The training mixture combines public biodiversity datasets, selected subsets of larger public image collections, and curated private Birder data sources. The goal of the corpus was broad visual coverage of the biological world rather than a fixed taxonomy, a supervised class list or a balanced benchmark-style dataset.

At a high level, the curation process follows the same concept as the Birder Vision Data Curation walkthrough for TreeOfLife-10M: input validation, near-duplicate removal, example-based filtering of non-natural images, quality/aesthetic filtering, and diversity-oriented sampling. The filtering process was designed to emphasize real photographs of living organisms and reduce illustrations, documents, maps, charts, diagrams, screenshots, lab imagery, medical imagery and other content outside the intended domain.

The final corpus includes broad biological coverage across plants, fungi, insect

From the published model card. Full card on the HuggingFace links in the sidebar.

How it works

How embedding models work

Your textsentence / documentEncodermaps meaningVectorlist of numbersAn embedding model turns text into a vector, so similar meanings sit close together — the basis of search and RAG.

Using it via the API

Call it like any OpenAI endpoint

Once AxForge deploys vit-reg4-so150m-p14-ls-dino-bio for you, it answers on the OpenAI-compatible API — the same base URL and keys as every other model. (vit-reg4-so150m-p14-ls-dino-bio below is illustrative; you get the exact model name on deployment.)

$ curl -sS https://api.axforge.ai/v1/embeddings \
  -H "Authorization: Bearer $AXFORGE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"vit-reg4-so150m-p14-ls-dino-bio","input":"text to embed"}'

Create an account — your API key is available in the console. 3M free tokens every 30 days with every new account.

© 2026 AxForge · EU-hosted AI infrastructure Pricing Docs Trust Privacy Terms