Correction Training improves diffusion sample quality without CFG
_onionesque · x · 2026-10-19
- Correction Training for Diffusion Models (donghoonahn et al.) significantly improves sample quality without classifier-free guidance (CFG).
- sedielem highlights it as related to nested looping at a fixed noise level, which fixes errors that would otherwise accumulate and reduces CFG reliance.
- Connects in spirit to data-space iteration, self-conditioning, deep equilibrium models, and fixed-point flows.
- Project page and technical thread available.
More from Research
- Doris Tsao's team invites open peer review on bioRxiv, bypassing journal timelines — anshulkundaje · 2026-10-19
- Amazon's Whole-Body Dexterity hits 90% success teaching Unitree G1 to haul boxes like humans — CyberRobooo · 2026-10-19
- ICML 2026 study: nearly half of LLM benchmarks are saturated — expert curation, not public data, decides longevity — evijit · 2026-10-19
- Nick Cammarata thread on the ramifications of mechanistic interpretability — spinozasrobot · 2026-10-19
- AI math debate: does traditional math culture block P/NP and Riemann progress? — _onionesque · 2026-10-19
- Decompressive cellular automata: pixels as 128-bit data, repeatedly 'decompressed' like brotli — zzznah · 2026-10-19