• Release gradio demo.
• Release technical report.
• Release training code.
• Release data processing code.

• 🔓 Fully Open-Source Texture Generation Pipeline

• 🧩 Compatible with Mainstream 3D Geometry Generation Methods

• 🎨 High-Fidelity, Geometry- and Image-Aligned PBR Texture Generation

• [2025-04-30] We have released our technical report!
• [2025-04-30] We have released model weights, data processing, trainning, inference scripts.

Installation guides can be found in INSTALL.md.

1. Prepare the input images and glbs in /data/test/imgs and /data/test/glbs.

2. Run the following command to infer:

   python infer_multiview.py --input_glb_dir ./data/test/glbs --input_img_dir ./data/test/imgs --output_dir ./outputs/test --geo_rotation 0
   

   Mesh generated from TripoSG, use --geo_rotation 0

   Mesh generated from Hunyuan-2, use --geo_rotation -90

   Mesh generated from TRELLIS，use --geo_rotation -90

   Mesh generated from Hi3dGen --geo_rotation -90

Extract PBR attributes from generated mul-view images.

python infer_pbr.py --mv_res_dir ./outputs/test

The extracted PBR maps will be saved in --mv_res_dir directory.

Run the following command to paint 3d model with multi-view generations:

python infer_paint.py --mv_res_dir ./outputs/test/mvpainter --output_dir ./results/test/mvpainter

If use PBR:

python infer_paint.py --mv_res_dir ./outputs/test/mvpainter --output_dir ./results/test/mvpainter --use_pbr

--mv_res_dir is the directory of the generated multi-view images.

python train.py --base=./configs/mvpainter-train-unet.yaml --gpus 0 --num_nodes 32 --logdir=./logs

python train.py --base=./configs/mvpainter-train-controlnet.yaml --gpus 0 --num_nodes 32 --logdir=./logs

Our pbr model is improved based on IDArb. The specific improvements can be found in our technical report. Our training script is similar to IDArb:

accelerate launch --config_file configs/acc/8gpu.yaml train_pbr.py --config configs/train.yaml

Dataprocess guides can be found in here

Our human-aligned evaluation system can be found in here

If you found this repository helpful, please cite our report:

@article{shao2025mvpainter,
  title={MVPainter: Accurate and Detailed 3D Texture Generation via Multi-View Diffusion with Geometric Control},
  author={Shao, Mingqi and Xiong, Feng and Sun, Zhaoxu and Xu, Mu},
  journal={arXiv preprint arXiv:2505.12635},
  year={2025},
  url={https://arxiv.org/abs/2505.12635}
}

In this project we use parts of the implementations of the following works:

• Hunyuan3D-1
• Hunyuan3D-2
• IDArb
• MaterialAnything

We thank these work's contributors for open sourcing their research and exploration.