NI-Tex introduces a novel approach to non-isometric garment texture generation by utilizing a physically simulated dataset, 3D Garment Videos, which provides consistent geometry and material supervision across diverse deformations. The method employs Nano Banana for high-quality non-isometric image editing, enabling reliable cross-topology texture generation. Additionally, an iterative baking process guided by uncertainty-driven view selection merges multi-view predictions into seamless, production-ready PBR textures. This results in versatile, spatially aligned garment materials, advancing industry-level 3D garment design. 🙌🙌🙌

• 🚀 Inference Code
• 📥 Model Checkpoints
• 📂 Training Dataset
• 🏋️‍♂️ Training Code

1. 💻 Environments. (cuda 12.4 on H100/H200)

git clone --recursive https://github.com/SII-Hui/NI-Tex.git
cd NI-Tex
conda env create -f environment.yml
conda activate NI-Tex

pip install basicsr==1.4.2 gfpgan==1.3.8 realesrgan==0.3.0 --no-deps

pip install torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric -f https://data.pyg.org/whl/torch-2.5.1+cu124.html

pip install custom_rasterizer/.

2. 📥 Pretrained Model Weights.

You need to manually download the RealESRGAN weight to the ckpt folder using the following command:

wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth -P ckpt

Download the pre-trained weights from Hugging Face: NI-Tex and place them in the ./MODEL_CHECKPOINTS/ directory.

To get started quickly, place your custom data into a newly created folder under the asset/cases directory. Please ensure your files follow this specific naming convention:

• 🧵 Mesh file: Rename your provided mesh to mesh.glb.
• 🖼️ Image prompt: Rename your provided image prompt to image_prompt.png.

📁 Example Directory Structure:

asset/
└── cases/
    └── your_custom_folder/
        ├── mesh.glb
        └── image_prompt.png

⚙️ Configuration & Tuning:

• --orth_scale: Controls the Stage 1 object size, which directly affects Stage 2 multi-view generation. Fine-tune this value for optimal results.
• resume_from (in inference.yml): We recommend MODEL_CHECKPOINTS/step_100K.ckpt for most cases. Use step_200K.ckpt for extreme or challenging inputs.

cd NI-Tex
python inference.py --name "GeneratedMesh_shirt" --base cfgs/inference.yml --orth_scale 1.35 --output_dir InferenceResults/

Our dataset rendering pipeline is inspired by MaterialAnything, utilizing assets from Objaverse 🌍, Texverse 🌍, and BEDLAM 👕.

We edited the BEDLAM data using Nano Banana 🍌 and rendered this massive dataset via Blender Python across a cluster of 48 RTX 4090 GPUs 🖥️.

To support future research, we plan to open-source our entire training dataset at Hugging Face: NI-Tex Dataset(to our knowledge, the first of its kind). As data preparation takes time, we will release it in the following order:

• 1. Bedlam_edited_by_NanoBanana ✨
• 2. BEDLAM (Rendered) 👕
• 3. Texverse (Rendered) 🌍
• 4. Objaverse (Rendered) 🌍

python train.py --base cfgs/hunyuan-paint-pbr.yaml --name overfit --logdir training_logs/ --gpus 0,

We have intensively borrow codes and dataset from the following repositories. Many thanks to the authors for sharing.

• 🔗 Hunyuan3D-2.1
• 🔗 MaterialAnything
• 🔗 Objaverse
• 🔗 Texverse
• 🔗 Bedlam
• 🔗 4DDress

If you find this repository useful in your project, please cite the following work. :)

@article{shan2025ni,
  title={NI-Tex: Non-isometric Image-based Garment Texture Generation},
  author={Shan, Hui and Li, Ming and Yang, Haitao and Zheng, Kai and Zheng, Sizhe and Fu, Yanwei and Huang, Xiangru},
  journal={arXiv preprint arXiv:2511.18765},
  year={2025}
}