TINKER: Diffusion's Gift to 3D--Multi-View Consistent Editing From Sparse Inputs without Per-Scene Optimization

📣 News

• [2025-08-21] Paper Released!

🚀 Overview

📖 Description

Compared with traditional approaches that either rely on per-scene optimization to maintain consistency or require fine-tuning to generate multi-view consistent edited images for every individual scene, Tinker is able to generalize without any additional training. Starting from sparse inputs, it produces dense, multi-view consistent edited images for both object-level and scene-level edits, thereby enabling fast 3DGS editing.

🚩 Plan

• Data and Data Pipeline
• Source code of Scene Completion Model
• Pipeline of 3DGS editing using NeRFStudio

🎫 License

For academic use, this project is licensed under the 2-clause BSD License. For commercial use, please contact Chunhua Shen.

🖊️ Citation

If you find this work useful, please consider citing:

@article{zhao2025tinkerdiffusionsgift3dmultiview,
    title={Tinker: Diffusion's Gift to 3D--Multi-View Consistent Editing From Sparse Inputs without Per-Scene Optimization}, 
    author={Canyu Zhao and Xiaoman Li and Tianjian Feng and Zhiyue Zhao and Hao Chen and Chunhua Shen},
    year={2025},
    journal={arXiv preprint arXiv:2508.14811}, 
}