Reproduce “Training-Free Multi-Concept Image Editing” 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
Niki Foteinopoulou, Ignas Budvytis, Stephan Liwicki
Concept Distillation Sampling (CDS) introduces a training-free framework for zero-shot, multi-concept image editing, addressing the limitations of text-only guidance and unstable multi-LoRA composition. It improves semantic alignment and structural preservation in image edits, outperforming baselines on benchmarks like InstructPix2Pix and ComposLoRA.
Key insights
- Refining the distillation process with ordered timesteps and explicit regularization prevents gradient decay and ensures more coherent, stable edits by controlling the denoising trajectory.
- Integrating negative prompt guidance directly into the optimization loop effectively mitigates degenerate visual modes often caused by aggressive LoRA conditioning.
- Dynamic, patch-wise concept weighting based on the divergence of LoRA noise predictions from the base model allows for seamless, spatially isolated integration of multiple visual concepts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.