UV 기반 텍스처 전이를 통한 모델 전환

Mesh-agnostic UV texture transfer.

Setup

Create conda environment

conda env create -f environment.yml
conda activate uvtt

Install Pytorch3D

pip install "git+https://github.com/facebookresearch/pytorch3d.git@stable"

Usage

Compute diffusion 3D features

python get_features.py 'data/source/mesh'

This will compute the vertex features for each mesh within 'data/source/mesh' and save them in 'data/source/features'. The number of meshes within the folder is arbitrary. You can put the whole body mesh or separate them into head and body like I did. The prompt will be based on the name of the obj file (human_head.obj = 'human head').

The seed for the diffusion model is currently fixed (check diff3f/diffusion.py).

python get_features.py 'data/target/mesh'

Do the same for the target mesh.

The names 'source' and 'target' are arbitrary. Just match the following directory structure:

.
├── data/
│   ├── source/
│   │   ├── mesh/
│   │   └── features/
│   └── target/
│       ├── mesh/
│       └── features/

Transfer textures

python transfer_textures.py data/source data/target

Provided that you computed the diffusion 3D features, this code will transfer the source textures from the source UV to the target UV.

Evaluation

Creating ground truth texture output

While this method is mesh agnostic, creating ground truth target texture from the source only works when the mesh connectivity is identical (check the code for required inputs)

python utils/reproject_textures.py

Calculate Fréchet inception distance (FID)

python evaluate_textures.py data/target/textures output/source2target