Skip to content

README.md

Latest commit

 

History

History
42 lines (33 loc) · 1.27 KB

README.md

File metadata and controls

42 lines (33 loc) · 1.27 KB

(SIGGRAPH Asia 2025) TC-GS: A Faster and Flexible 3DGS Module Utilizing Tensor Cores (paper,slide)

TC-GS is a flexible and fast library which can accelerate the renderCUDA process of 3DGS with Tensor Cores. It can be easily installed with various 3DGS kernels.

This repo is an example applying Speedy-splat with TC-GS. We have also apply TC-GS on other acceleration kernels and achieving remarkable speedup.

The code and usage of TC-GS is in submodules/tcgs_speedy_rasterizer/tcgs.

TODO

  • Support Training with Tensor Cores
  • Utilizing Tensor Cores on preprocessCUDA

Installation

git clone https://github.com/DeepLink-org/3DGSTensorCore --recursive

Setup

conda env create --file environment.yml

Evaluate The trained model

export DATA=[your_data_path]
export SCENE=[your_scene_name]
export CKPT=[your_checkpoint_path]

# export CUDA_VISIBLE_DEVICES=0

python render.py \
    -s ${DATA}/${SCENE}/ \
    -m ${CKPT}/${SCENE}/ \
    --eval

or simply use the script

bash eval.sh

Result

The result is evaluated on NVIDIA A800