SNaP: one-step posterior sampler hits 12ms per sample, 30-2250x faster than iterative
prof_kamilov · x · 2026-09-29
- SNaP is a one-step MeanFlow posterior sampler for linear inverse problems. Its key idea is a measurement-adapted source Gaussian whose mean and anisotropic covariance come from the measurement operator, observation, and noise level, anchoring well-measured directions while preserving variation elsewhere; the exact conditional flow transports this source to the true posterior.
- Performance: 1 NFE per posterior sample (one source solve plus one network evaluation), 30x+ faster than iterative samplers on CelebA deblurring, 12ms per 128x128 sample, and up to 2250x speedups claimed.
- Tested on natural image restoration and noisy multi-coil brain MRI, where a single evaluation yields competitive reconstructions; averaging M=4/16/100 draws trades fine detail for lower pixel error. Code and project page are public.
Related event: SNaP: One-Step Posterior Sampling Over 30x Faster Than Iterative Methods(3 posts)→
More from Research
- Economist John Horton: keep AI-generated papers out of human venues, publish on GitHub — soumitrashukla9 · 2026-09-29
- Harvard's Clinical Informatics Lecture Series hosts "AI and Mental Health" with Dr. John Torous — zakkohane · 2026-09-29
- HCOMP 2026 study: imagery scales skew annotator and reviewer performance — windx0303 · 2026-09-29
- EleutherAI open-sources Bergson, a scalable data attribution library for LLMs, with an EMNLP 2026 oral paper — BlancheMinerva · 2026-09-29
- Bergson data attribution library released with first public MAGIC implementation — BlancheMinerva · 2026-09-29
- EleutherAI tests whether EK-FAC data attribution can block subliminal learning — results are 'solidly mid' — BlancheMinerva · 2026-09-29