🐟 TaxaDiffusion: Progressively Trained Diffusion Model for Fine-Grained Species Generation

🟢 Accepted to ICCV 2025

---

🧑‍💻 Authors

Amin Karimi Monsefi, Mridul Khurana, Rajiv Ramnath, Anuj Karpatne, Wei-Lun Chao, Cheng Zhang

---

✨ Highlights

TaxaDiffusion is a novel framework that tackles the challenge of generating fine-grained species-level images using diffusion models. Biological species often differ by subtle visual traits, making it difficult for standard generative models to capture and preserve these distinctions. To address this, TaxaDiffusion leverages the hierarchical nature of biological taxonomy to progressively condition and guide the generation process.

We propose a stage-wise training strategy, where the model is first trained on coarse taxonomic categories (e.g., Class or Order) and is gradually fine-tuned on finer labels (e.g., Genus and Species). This allows the diffusion model to incrementally learn from general visual semantics to subtle, species-specific cues.

Key contributions:

• 🧬 Taxonomy-aware progressive training: We introduce a novel multi-stage framework that conditions diffusion generation on biological hierarchy, improving structure and diversity.
• 🔁 Stage-wise refinement of generations: Each stage refines the model’s output by incorporating finer taxonomic labels, resulting in more accurate and biologically coherent species synthesis.
• 📊 Comprehensive evaluation on three datasets:
  • FishNet – 17,000+ fish species with high inter-class similarity.
  • BIOSCAN-1M – Microscopic images of 8,355 insect species.
  • iNaturalist – Diverse set spanning 10,000 plant and animal species.
• 🏆 State-of-the-art results:
  • Achieves superior FID and LPIPS scores for image quality.
  • Improves BioCLIP-based alignment between species text labels and generated images.
  • Demonstrates strong generalization across single-group and mixed-species datasets.

Progressive training from high-level taxonomy to fine-grained species generation.

---

📦 Installation

Download the dataset and change the config with the path of the dataset:

#FishNet Dataset
https://fishnet-2023.github.io/

#BIOSCAN-1M Dataset
https://github.com/bioscan-ml/BIOSCAN-1M

# iNaturalist Dataset
https://www.tensorflow.org/datasets/catalog/i_naturalist2021

Clone the repository and set up the environment:

git clone https://github.com/aminK8/TaxaDiffusion.git
cd TaxaDiffusion

conda env create -f environment.yml
conda activate taxa_diffusion

Train

bash job_training.sh

Inference

bash job_inference.sh

Citation

If you liked our paper, please consider citing it

@article{monsefi2025taxadiffusion,
  title={TaxaDiffusion: Progressively Trained Diffusion Model for Fine-Grained Species Generation},
  author={Monsefi, Amin Karimi and Khurana, Mridul and Ramnath, Rajiv and Karpatne, Anuj and Chao, Wei-Lun and Zhang, Cheng},
  journal={arXiv preprint arXiv:2506.01923},
  year={2025}
}