Skip to content

Latest commit

 

History

History
49 lines (36 loc) · 2.33 KB

README.md

File metadata and controls

49 lines (36 loc) · 2.33 KB

CLIP-based Neural Neighbor Style Transfer for 3D Assets

This repository contains the code for our paper, CLIP-based Neural Neighbor Style Transfer for 3D Assets.

Dependencies

Before you install make sure you have CUDA capable devices. Without such devices, the code will not run.

Start with installing the dependencies:

pip install -q ftfy regex tqdm
pip install -q git+https://github.com/openai/CLIP.git
pip install -q ninja imageio PyOpenGL glfw xatlas gdown
pip install -q git+https://github.com/NVlabs/nvdiffrast/
pip install -q kmeans-pytorch
pip install -q --global-option="--no-networks" git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch

How to run

After you have everything installed, running is pretty simple and straight forward.

python optimize.py \
    --style-image ./assets/starry.jpg \
    --input-mesh ./assets/spot.obj \
    --output-path /tmp \
    --batch-size 4 \
    --iteration 15000 \
    --texture-res 1024 \
    --render-res 512 \
    --scale 1.5

In order to run in multi-GPU setting, please setup slurm or proper environment variables.

After complete optimization, you will get the stylized image.

Licenses

All the files inside the ./rasterizer folder is under NVIDIA's license. The original source code for files under ./rasterzier can be found in this link . I copied these files from nvdiffmodeling to accommodate changes according to updated version from nvdiffrec.

However, everything else is MIT license. Since the code requires the using nvdiffrast, it might be useful to get proper license for commercial purpose.

Asides from that feel free to use this on your personal projects and research. The NVIDIA open source license allows all non commercial work.

Assests