A new preprint on arXiv introduces StimFlow, a method for synthesizing brain activity from images. The authors frame the task as a generative problem, arguing that fMRI responses vary naturally across trials and therefore require a model that can produce a distribution of possible responses rather than a single deterministic output.

StimFlow is built on flow matching, a class of generative models, and is conditioned on the stimulus image. The abstract notes that prior generative encoders often operate in compressed latent spaces and condition predominantly on semantic information, implying that StimFlow aims to address limitations of those approaches—though the full details are not available in the abstract.

The paper is a preprint and has not yet been peer-reviewed. As the abstract is truncated, the reported results and specific architectural choices are not fully described here.