The official repo for "Texture-GS: Disentangling the Geometry and Texture for 3D Gaussian Splatting Editing"

[02/7/2024] Our paper is accepted by ECCV 2024. Welcome to discuss with us in MiCo Milano.

[28/5/2024] The code is released, including training code, viewer and pretrained models.

[18/3/2024] The project page is created.

[18/3/2024] The official repo is initialized.

3D Gaussian splatting, emerging as a groundbreaking approach, has drawn increasing attention for its capabilities of high-fidelity reconstruction and real-time rendering. However, it couples the appearance and geometry of the scene within the Gaussian attributes, which hinders the flexibility of editing operations, such as texture swapping. To address this issue, we propose a novel approach, namely Texture-GS, to disentangle the appearance from the geometry by representing it as a 2D texture mapped onto the 3D surface, thereby facilitating appearance editing. Technically, the disentanglement is achieved by our proposed texture mapping module, which consists of a UV mapping MLP to learn the UV coordinates for the 3D Gaussian centers, a local Taylor expansion of the MLP to efficiently approximate the UV coordinates for the ray-Gaussian intersections, and a learnable texture to capture the fine-grained appearance. Extensive experiments on the DTU dataset demonstrate that our method not only facilitates high-fidelity appearance editing but also achieves real-time rendering on consumer-level devices, e.g. a single RTX 2080 Ti GPU.

• Release the demo page and more video results.
• Release the code.
• Add more explanations.

Our experiements are conducted on the platform with CUDA=11.7, torch==1.13.1 and pytorch3d==0.7.6. However, other version should also be compatible.

Other dependencies have been listed in the requirements.txt.

DTU dataset: We follow Relightable 3D Gaussian to prepare DTU data, which contains a foreground mask and a psedo-ground truth normal maps for each scenes. Download the zip file of preprocessed DTU data here and then create a soft link to data/dtu/.

Any obj file: To create your synthetic dataset from obj files, such as Objaverse or Google Scanned Objects, we provide a preprocess script to render NeRF synthetic format dataset. We adopt blender=3.6.11 to render images, with the params elevs and num_renders to speficy the elevation angles of camera poses and the number of images per elevation angles, respectively.

cd scripts
blender -noaudio --background --python render_obj_file.py -- \
    --object_path path/to/your/obj/file \
    --output_dir path/to/your/dataset \
    --num_renders 12 \
    --elevs -30 0 30 60

⚠️ Limitations ⚠️: We define the texture space as a unit sphere, which is ill-suited to represent multiple objects, objects with complex geometry(e.g. NeRF Synthetic chair, lego) or outdoor scenes.

We provide several pretrained models on DTU dataset and other object-level datasets here to illustrate the appearance editing abilities of Texture-GS, especially global texture swapping. First, download the pretrained models and save them to the direction pretrained/. Then, run the following script to change the original texture with our provided creative texture images in assets/. Remember to modify data_root_dir in the configure file texture_gaussian3d.yaml with the corresponding model weights (remove --load_texture_from to visualize the original view synthesis results).

python retexture.py configs/texture_gaussian3d.yaml \
    --resume_from pretrained/dtu118.pth \
    --load_texture_from assets/textures/mosaic.png

The rendered images of DTU 118 with mosaic texture are stored to output/texture_gaussian3d/tex_[timestamp0]. Here're two examples of rendered images

If you wanna visualize the learned texture images, run the following line to extract the texture image

python extract_texture.py configs/texture_gaussian3d.yaml \
    --resume_from pretrained/dtu118.pth \
    --save_path pretrained/dtu118_tex.png

Here's the corresponding texture image of pretrained DTU 114

We provide a real-time Open-GL based renderer for Texture-GS, which is built on source code shared by Tiny Gaussian Splatting Viewer. Run the following lines (remove --load_texture_from to visualize the original view synthesis results)

python viewer.py configs/texture_gaussian3d.yaml \
    --resume_from pretrained/dtu118.pth \
    --load_texture_from assets/textures/mosaic.png

Our training stage is composed of three steps, including geometry reconstruction, UV mapping learning and texture reconstruction. First, run the training scripts train.py and specify the config file configs/gaussian3d_base.yaml to obtain the initial 3D Gaussian-based geometry, use CUDA_VISIBLE_DEVICES to specify the GPU number

python train.py configs/gaussian3d_base.yaml

The checkpoints, point clouds of 3D-GS, config files and tensorboard event file during training are saved to output/gaussian3d_base/[timestamp1]/. Then, extract a point cloud from the checkpoints to obtain a psedo ground truth point cloud that roughly covered the underlying surface for UV mapping learning. Run the scripts extract_pcd.py and specify path to the checkpoint.

python extract_pcd.py configs/gaussian3d_base.yaml \
    --resume_from output/gaussian3d_base/[timestamp1]/checkpoints/30000.pth \
    --save_path output/gaussian3d_base/[timestamp1]/pcd.npy

The extracted point cloud pcd.npy as well as a ply file pcd.ply for visualization on VSCode (we recommend vscode-3d-preview extension) is generated in the save path. After obtaining the psedo GT, we can learn a UV mapping corresponding to the geometry with Chamfer Distance loss. Notably, you need to modify the following lines in the config file configs/uv_map.yaml

    init_from: output/gaussian3d_base/[timestamp1]/checkpoints/30000.pth
    pcd_load_from: output/gaussian3d_base/[timestamp1]/pcd.npy

Then, run the following training script. Notable, the PSNR L1 and SSIM metrics output by the scripts are meaningless here, because we only concern about the geometry of 3D Gaussians.

python train.py configs/uv_map.yaml

The checkpoints, config files and tensorboard event file during training are saved to output/uv_map/[timestamp2]/. We also visualize the inverse UV mapping process in output/uv_map/[timestamp2]/pcds/ by randomly sampling points on UV space and then projecting them back to 3D space. Finally, we modify the following lines in configs/texture_gaussian3d.yaml to specify the initial path for 3D Gaussian and UV mapping, and then train our Texture-GS

    init_from: output/gaussian3d_base/[timestamp1]/checkpoints/30000.pth
    init_uv_map_from: output/uv_map/[timestamp2]/checkpoints/15000.pth

python train.py configs/texture_gaussian3d.yaml

We take the training process of DTU scan 118 as example. If you want to train Texture-GS on other scenes, such as DTU scan 122, just modify data_root_dir in all config files. We train our model with 800*600 resolution images on the DTU dataset to compare with previous works. However, the original resolution 1600*1200 is also supported by setting the following line resolution: -1.

Due to the lack of psedo ground-truth normal maps, we train Texture-GS on object derived dataset with configure files configs/gaussian3d_base_object.yaml configs/uv_map_object.yaml and configs/texture_gaussian3d_object.yaml, which are slightly different from DTU configure files. Here's a simple example of STACKING_BEAR from GSO dataset.

If you find this code useful for your research, please consider citing:

@misc{xu2024texturegs,
    title={Texture-GS: Disentangling the Geometry and Texture for 3D Gaussian Splatting Editing}, 
    author={Tian-Xing Xu and Wenbo Hu and Yu-Kun Lai and Ying Shan and Song-Hai Zhang},
    year={2024},
    eprint={2403.10050},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

This project is built on source codes shared by 3DGS and diff-gaussian-rasterization.