PRISM Framework for Controllable Unpaired Image Translation
August 7, 2026
PRISM introduces a GAN-free flow-matching framework that uses distribution-gated feature selection for image-to-image translation. By replacing global noise with a learned per-feature gate based on distance to target distributions, it enables precise control over which content is preserved and which appearance is modified.
HOW THIS AFFECTS YOU
●
researcherThis approach moves beyond global guidance by utilizing spatial priors for feature-level control in flow matching.
●
designerYou gain more granular control over specific elements of an image during translation tasks.