Test-time scaling of diffusion models using sequential Monte Carlo (SMC) has emerged as a promising strategy for generating samples that satisfy user-specified rewards, such as human preference scores. The approach allows models to be steered at inference time without retraining, making it attractive for aligning text-to-image outputs with qualitative goals.
The new arXiv preprint applies this idea to flow matching models, a class of generative models closely related to diffusion models. The authors introduce a noise-rejuvenated SMC technique, which appears designed to overcome shortcomings of the standard SMC framework. The abstract notes that the standard framework "typically" encounters issues, though the specific details are cut off in the available text.
Given the truncated abstract, the exact mechanics of the noise rejuvenation and the reported results are not fully described here. What is clear is that the paper positions noise-rejuvenated SMC as a way to make test-time scaling more effective for text-to-image flow matching, building on prior work in diffusion-based SMC sampling.