DealMaTe: Multi-Dimensional Material Transfer via Diffusion Transformer

These are the pre-trained Shader LoRA weights for DealMaTe — an efficient, text-free material transfer framework built on FLUX.1 Diffusion Transformer.

Given a material reference image and a target object image, DealMaTe faithfully transfers the material appearance onto the object while preserving its 3D geometry, lighting, and surface structure.

Paper: ACM Transactions on Graphics (TOG), 2026

Model Weights

Three Shader LoRA adapters, each encoding a distinct intrinsic component for the FLUX.1 model:

File	Size	Description
depth.safetensors	286 MB	Depth LoRA — encodes 3D spatial structure
normal.safetensors	286 MB	Normal LoRA — captures surface curvature
lighting.safetensors	286 MB	Lighting LoRA — models illumination direction and intensity

Usage

These weights are designed for the DealMaTe inference pipeline. Clone the repo and place the three .safetensors files in the lora/ directory:

git clone https://github.com/haha-lisa/DealMaTe.git
cd DealMaTe

# Download from Hugging Face
pip install huggingface_hub
huggingface-cli download lisalisalisa/DealMaTe depth.safetensors normal.safetensors lighting.safetensors --local-dir ./lora

Then run inference:

python inference.py \
    --material_path  examples/inputs/material.png \
    --content_path   examples/inputs/content.jpg \
    --mask_path      examples/inputs/mask.png \
    --depth_path     examples/inputs/depth.png \
    --normal_path    examples/inputs/normal.png \
    --lighting_path  examples/inputs/lighting.png \
    --output_path    outputs/result.png \
    --lora_path      ./lora

Required base models (auto-downloaded):

• black-forest-labs/FLUX.1-dev
• black-forest-labs/FLUX.1-Redux-dev

Requirements: Python 3.10+, CUDA 11.8+, ~40 GB GPU VRAM (A100 recommended).

Citation

@article{huang2026dealmate,
  title     = {DealMaTe: Multi-Dimensional Material Transfer via Diffusion Transformer},
  author    = {Huang, Nisha and Lin, Yizhou and Guo, Jie and Li, Xiu and Lee, Tong-Yee and Yu, Zitong},
  journal   = {ACM Transactions on Graphics},
  year      = {2026},
  publisher = {ACM}
}

@inproceedings{huang2025mate,
  title  = {MaTe: Images Are All You Need for Material Transfer via Diffusion Transformer},
  author = {Huang, Nisha and Liu, Henglin and Lin, Yizhou and Huang, Kaer and Chen, Chubin and Guo, Jie and Lee, Tong-yee and Li, Xiu},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages  = {15117--15126},
  year   = {2025}
}

License

LoRA weights are released under the MIT License. The base FLUX.1 model is subject to its own license.