Skip to content

Repository files navigation

Website arXiv License

Video

Demo

Installation & Usage

Environment Setup

Option 1. (Tested on CUDA 12.8) Setup with conda.

conda install -c conda-forge cuda-toolkit=12.8
conda create -n lumitex python=3.11
conda activate lumitex

pip install torch==2.7.0 torchvision==0.22.0 torchaudio==2.7.0 --index-url https://download.pytorch.org/whl/cu128
pip install -r requirements.txt

# install Hunyuan3D-2.1 renderer
cd Renderer/custom_rasterizer
pip install --no-build-isolation -e .
cd ../DifferentiableRenderer
bash compile_mesh_painter.sh
cd ../..

Option 2. (Tested on CUDA 12.8) Setup with pixi. Make sure you have pixi installed.

pixi shell
pixi run post-install

Model Weights

Please download the pre-trained model weights from Huggingface (FLUX-MV-Shaded), and the lightweight LoRA weights (trained on data Step1X-3D).

Assume the model is put on the ckpt folder in inference.py by default.

Inference

Data: Structure your data (reference image and mesh) as in the tests folder: ref.png, *.obj/glb.

Then run the command as below, the pipeline generates and inpaints basecolor by default (more details see inference.py).

python inference.py --folder tests/case_1

Acceleration

Pipeline acceleration techniques like torch.compile, cache-dit, flash-attn3 are compatible with our pipeline.

Results

TODO

  • Evaluation code.
  • Inference pipeline and pretrained model.
  • Data processing and infra code.

Acknowledgement

We would like to express our gratitude to the authors of the following projects for their efforts and open contributions: Diffusers, FLUX.1-dev, Huggingface, UniTEX, LVSM, LaCT, Hunyuan3D-2.1, Hunyuan3D-2.5.

License

This repository is licensed under the Apache License 2.0. This project also uses third-party components with non-commercial license: FLUX model, Tencent Hunyuan3D-2.1 rendering code. These components are subject to their respective licenses and may restrict commercial use. Users are responsible for complying with all applicable licenses.

BibTeX

@inproceedings{
  lumitex2026,
  title={LumiTex: Towards High-Fidelity {PBR} Texture Generation with Illumination Context},
  author={Jingzhi Bao and Hongze Chen and Lingting Zhu and Chenyu Liu and Runze Zhang and keyang luo and Zeyu HU and Weikai Chen and Yingda Yin and Xin Wang and Zehong Lin and Jun Zhang and Xiaoguang Han},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026},
  url={https://openreview.net/forum?id=CDwG0Bebfo}
}

About

[ICLR 2026] LumiTex: Towards High-Fidelity PBR Texture Generation with Illumination Context

Resources

Stars

70 stars

Watchers

8 watching

Forks

Releases

Packages

Contributors

Languages