Developed noise-rejuvenated SMC (NR-SMC) to apply inference-time scaling to deterministic flow matching models.
Noise-rejuvenated SMC enables test-time scaling for flow matching image generation
New method injects stochasticity only at resampling steps, preserving generation quality while steering outputs toward user-specified rewards
Academic
Wenyi Lian · Zheng Zhao · Thomas B. Sch{ö}n
Uppsala University · Linköping University
Research Digest··2 min read
The authors introduce noise-rejuvenated sequential Monte Carlo (SMC), an inference-time scaling method for text-to-image flow matching models.
§