PRISM: Distribution-Gated Flow Matching for Controllable Unpaired Image Translation
PRISM replaces one-shot global guidance with a learned feature-level gate for deciding what an image translator should alter.
The paper says the gate is based on how far each source feature sits from the target distribution, freeing mismatched features while preserving target-consistent ones. That same gate controls both the corrupted latent initialization and the timing of flow-matching transport during ODE integration. The authors report tests on five natural and biomedical benchmarks, with PRISM leading FID and KID on four and staying competitive on the fifth. In histopathology, they say it also kept nuclei counts closest to the ideal balance between realism and structure preservation. ArXiv · AI/CL/LG's note
The paper says the gate is based on how far each source feature sits from the target distribution, freeing mismatched features while preserving target-consistent ones. That same gate controls both the corrupted latent initialization and the timing of flow-matching transport during ODE integration. The authors report tests on five natural and biomedical benchmarks, with PRISM leading FID and KID on four and staying competitive on the fifth. In histopathology, they say it also kept nuclei counts closest to the ideal balance between realism and structure preservation. ArXiv · AI/CL/LG's note
score 4