Interpretable Procedural Material Graph Generation via Diffusion Models from Reference Images

This is the official implementation of the paper Interpretable Procedural Material Graph Generation via Diffusion Models from Reference Images

Requirements

git clone https://github.com/1278323067/procedural_texture.git
cd procedural_texture

conda env create -f timm_texture.yaml
conda activate timm_texture

Dataset

The dataset is available here , which includes images and JSON files (containing the parameter attributes of procedural nodes corresponding to the images).

Usage

checkpoint

The adpater checkpoint file is available here . After downloading, place it in the \ckpt folder.

The Vit classifier ckeckpoint file is available here

inference

The inference file is in experiment/sampler.py

Note that:

• --adapter_path: The adpater checkpoint file path '.ckpt/img_emb_model'.
• --cls_model_pretrained: The Vit classifier ckeckpoint file path.
• --node_type_path: node_type.pkl
• --image_input: image file path

 
python experiment/sampler.py

After running this , the results are in experiment/output/

You will get results like this

run ./diffmat_1/test/test_hybrid_optimizer.py to get the optimization parameters, more details see diffmat.

These generated nodes and parameters can later be used in Substance Designer