Reproduce “ColorEdit: Training-free Image-Guided Color editing with diffusion model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xingxi Yin, Zhi Li, Jingfeng Zhang, Chenglin Li, Yin Zhang
ColorEdit, developed by researchers at Zhejiang University, introduces a training-free, image-guided framework for object-level color editing with diffusion models, addressing limitations like "attention leakage" and "attribute collision" in text-guided approaches. The method achieves state-of-the-art performance in color modification, object structure, and background preservation, demonstrating superior scores on a newly introduced COLORBENCH dataset, including the lowest LPIPS_bg (0.016) for background preservation.
Key insights
- An object's shape, contour, and texture are primarily established in the U-Net decoder (1st, 2nd, and 3rd CrossAttnUpBlocks) during the early stages of the denoising process.
- Failures in text-guided color editing arise from "cross-attention leakage" and "attribute collision" within cross-attention maps, leading to imprecise or ineffective color changes.
- Modifying Value matrices in cross-attention layers yields more stable and accurate color changes compared to altering Key matrices, which can introduce structural distortions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.