Ecosystem Roadmap

Learn about the future directions of the BioCLIP ecosystem, including plans for tools, models, data, and community building initiatives. This roadmap represents our vision for advancing biodiversity science through innovative machine learning applications and collaborative efforts.

Browse Areas of Focus Detailed Future Directions

Roadmap topics for 2026 and beyond

Software

Plans for new and improved BioCLIP and TreeOfLife software and tools

Go to Software Development →

Community

Plans for future community building and engagement initiatives

Go to Community Development →

Models

Plans for future BioCLIP model development

Go to model development →

Data

Plans for future TreeOfLife data development

Go to Data Development →
Software

Software Development

We have developed a collection of open-source software tools to facilitate the use of BioCLIP models and TreeOfLife data processing and development. The next steps for software development include improving the interpretability of model inference, enhancing the ease of use of our tools, and improving the demos for our software.

  • Interpretability: Incorporate inference interpretability methods, including Finer-CAM, Prompt_CAM, and SAEs.
  • Easy usage: Mobile app, geo-fencing, detection-classification pipeline (multi-species), interactive patch masking (multi-species), selectable taxonomy level.
  • TaxonoPy : Improved coverage and handling of common names.
  • Improved Demos: Hosted Image Embedding Explorer and BioCLIP Image Search, with images, not URLs. Added demos for self-hosted apps.

Existing Software
Software Plans
Community

Community Development

The BioCLIP Ecosystem consists of a collection of open-source products (software, models, data, and apps) designed to facilitate biological research. We aim to foster a vibrant and inclusive community through various initiatives, including:

  • BioCLIP Helpdesk: Centralized and community helpdesk following the open-source model of maintainer and community feedback and assistance.
  • WildLabs Group: Establish a group in the global WildLabs community for collaboration and knowledge sharing.
  • Community Forum: Shared space for discussions and collaboration.
  • Community Highlights: Showcase community achievements and contributions. Include a clear workflow for submitting projects to spotlight.
  • Community Contributions: Centralized platform for community members to find issues and projects open to contribution.
  • Leaderboard: Maintain a community leaderboard ranking BioCLIP and associated models on various biological benchmarks. Include method for submitting and evaluating models for inclusion on the leaderboard. This would include links to models and source code, in addition to the rankings.

Existing Community Initiatives
Community Development
Models

Models Development

We have demonstrated emergent properties through scaling up and enriched biological semantic understanding through synthesized captions. The next steps for BioCLIP model development include hierarchical consistency, accessibility improvements, efficient adaptation, and applicability expansion.

  • Hierarchical Consistency: With previous models achieving accurate species classification under the guidance of taxonomic labels, we aim to further leverage this structure to improve hierarchical consistency.
  • Accessibility Improvements: BioCLIP models now produce accurate species recognition on GPU-accelerated hardware. We want to improve the scalability and accessibility of our models for realtime and edge device applications, such as camera traps, mobile apps, and web applications.
  • Efficient Adaptation: BioCLIP models were optimized for classification at the scale of the Tree of Life using TreeOfLife-200M. However, for downstream applications involving specific data domains or species coverage, model adaptation can further improve accuracy and efficiency. We aim to investigate best practices for adapting BioCLIP models across diverse application scenarios.
  • Applicability Expansion: BioCLIP models produce image-aligned taxonomic embeddings that enable applications beyond recognition, including fine-grained species generation (e.g., TaxaAdapter). We aim to explore novel applications of BioCLIP, inspired by the broad applicability of CLIP.

Existing Models
Models in Development
Data

Data Development

Starting from TreeOfLife-10M, we have explored scaling up in TreeOfLife-200M and enriching captions in TreeOfLife-10M-Captions. The next steps for data development include improving the quality of captions, balancing the data distribution, and resolving species common names more consistently.

  • Caption Quality: The synthesized descriptive captions, grounded by Wikipedia-derived contexts, have helped BioCAP acquire better understanding of biological semantics. We aim to improve the captions through more comprehensive knowledge integration and further expand the captioning to larger scales.
  • Data Distribution Balancing: Biological visual datasets often exhibit long-tailed distributions, with a few common species dominating the data and many rare species underrepresented. We aim to explore methods for balancing the data distribution, such as data selection, and investigate the influence of data on model performance.
  • Species Name Resolution: TaxonoPy provides a solid initial framework for resolving species names across different data sources. We aim to further study the optimal strategy of species name resolution and the impact of name ambiguity on model performance.
Existing Datasets and Benchmarks
Data in Development