Generative Zoo

Tomasz Niewiadomski, Anastasios Yiannakidis, Hanz Cuevas-Velasquez, Soubhik Sanyal,
Michael J. Black, Silvia Zuffi, Peter Kulits

Project Page • Demo

---

📖 Overview

Generative Zoo (GenZoo) provides a scalable pipeline for generating realistic 3D animal pose-and-shape training data.
Models trained exclusively on GenZoo data achieve state-of-the-art performance on real-world 3D animal pose and shape estimation benchmarks.

---

⚙️ Installation

1. Install uv.
2. Clone the repository:

   git clone https://github.com/x-tomasz/genzoo
   cd genzoo

3. Sync the uv environment:

uv sync --locked

---

📂 Model Weights & Configs

1. Register on the project website, download the following files, and place them in ./data/:

   • genzoo_1M.ckpt
   • smal_plus.pkl
   • animal_attributes.npz
   • dog_poses.npz

2. Download the following external dependencies into ./data/:

   • submission_animal_realnvp_mask_pred_net_6000.pth
   • vitpose_backbone.pth

---

🚀 Inference

Run inference on cropped input images:

uv run python inference.py ./eg_input --render

---

🛠️ Data Generation Pipeline

Start the GenZoo pipeline notebook:

uv run jupyter notebook genzoo_pipeline.ipynb

On a headless server:

uv run jupyter notebook genzoo_pipeline.ipynb --no-browser --port=8888 --ip=0.0.0.0

---

🎓 Training

1. Download and extract Pascal VOC (≈2 GB) under ./data/ for synthetic augmentation (see isarandi/synthetic-occlusion).

2. Download the GenZoo training datasets:

   • v1: weak perspective camera
   • v2: full perspective camera

(Uncomment ## Full perspective code in hmr2/datasets/image_dataset.py to use full perspective camera GT)

3. Prepare WebDataset tars with your chosen paths:

uv run python prep_tars.py

4. Choose test set, and specify the path in hmr2/configs/datasets_tar.yaml, COCO-VAL field. (Default is Animal3D (real) or GenZoo-Felidae (synthetic))

5. Run training:

uv run python train_genzoo.py exp_name=your_exp_name \
  data=mix_all experiment=hmr_vit_transformer trainer=gpu launcher=local

---

🙏 Acknowledgments

This project builds on and adapts parts of:

• 4D-Humans
• AWOL
• SMAL
• Synthetic Occlusion

Special thanks to all GenZoo co-authors for their contributions and support, and in particular to Peter Kulits for supervision and guidance.

---

🎤 ICCV 2025 Presentation 🌺🌴🌊

GenZoo will be presented as a poster at ICCV 2025 in Hawaii:

• Poster ID: #788
• Session: Tue, 21 Oct 2025 — 6:15 p.m. to 8:15 p.m. PDT
• Authors: Tomasz Niewiadomski, Anastasios Yiannakidis, Hanz Cuevas Velasquez, Soubhik Sanyal, Michael J. Black, Silvia Zuffi, Peter Kulits

---

📝 Citation

If you use GenZoo in your research, please cite:

@inproceedings{niewiadomski2025ICCV,
  author    = {Niewiadomski, Tomasz and Yiannakidis, Anastasios and Cuevas-Velasquez, Hanz and Sanyal, Soubhik and Black, Michael J. and Zuffi, Silvia and Kulits, Peter},
  title     = {Generative Zoo},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year      = {2025}
}

---

📜 License

We build off the 4D-Humans codebase to perform our experiments. As such, inherited code falls under the original MIT license. Additions and modifications are released under a different license in accordance with institute requirements which has been prepended to LICENSE.md.