Reproduce “Adaptive Inference-Time Scaling via Early-Step Latent Verification for 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
Yue Yu, Yang Jiao, Jiayu Wang, Qi Dai, Jingjing Chen
VeriLatent screens image-editing candidates using early latent changes and predicted edit-region masks, rather than decoding every trial for inspection. After retaining promising candidates, it fully renders them and assesses final edits. On GEdit-Bench-EN with FLUX.1 Kontext and a maximum budget of 32, it scores 7.637 for semantic consistency and 7.082 overall versus ADE-CoT’s 7.225 and 6.695, while perceptual quality is lower.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.