SONIC Update: Fourier-Optimized Noise
kwangmoo_yi · x · 2026-07-10
This post shares an update on the ECCV'26 paper SONIC (Spectral Optimization of Noise for Inpainting with Consistency). The authors added a user study and further experiments, revealing that their method achieved a 90% preference rate among users.
The core focus remains on the paper's methodology: initial noise significantly impacts image inpainting results, and the authors propose optimizing this noise in the Fourier space via linearization, without needing to backpropagate through the denoiser.
More from Research
- Anthropic masterclass spotlights how to build and observe AI agents — _jaydeepkarale · 2026-07-21
- NeurIPS 2026 workshop calls papers on on-device intelligence — YiMaTweets · 2026-07-21
- AI Security Institute says every tested model tried to cheat in cyber evaluations — connoraxiotes · 2026-07-21
- AI companies are buying old books to avoid training on AI-generated slop — CackleRooster · 2026-07-21
- Sakana says multiple diffusion models plus MCTS beat test-time scaling on coding and math — SakanaAILabs · 2026-07-21
- Soofi S 30B-A3B releases a full pretraining report and claims open-model leads in English and German — abursuc · 2026-07-21