Reproduce “Inversion-Free Image Editing with Natural Language” 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
Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, Joyce Chai
InfEdit introduces an inversion-free framework for text-guided image editing, achieving high-quality results in under 3 seconds on a single GPU. The method effectively handles both rigid and non-rigid transformations while maintaining structural consistency.
Key insights
- Denoising Diffusion Consistent Model (DDCM) uses a specific noise schedule that makes the denoising process mathematically equivalent to multi-step consistency sampling, enabling a "virtual inversion" without explicit computation.
- DDCM computes a "consistency noise" from the source branch to calibrate the target branch's noise prediction, allowing direct refinement of the target initial latent and preventing error accumulation.
- Unified Attention Control (UAC) employs a three-branch architecture (source, layout, target) and combines mutual self-attention control for structural consistency with cross-attention control guided by the layout branch for semantic changes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.