PRISM: GAN-Free Flow Matching for Controllable Unpaired Image Translation
A new study introduces PRISM, a flow-matching framework designed for unpaired image-to-image translation without relying on GANs. This novel method replaces the usual global noise or guidance metrics with a learned gate tailored for each feature, improving accuracy in deciding which parts of an image to alter or keep. The gate's spatial prior is informed by the standardized distance of source features from the target feature distribution, allowing it to ignore far-off features while keeping relevant ones. It also oversees initialization by integrating the actual source latent with a specific corruption and regulates transport timing during Ordinary Differential Equation (ODE) integration. The content-anchored corruption (AdaIN) helps maintain structural integrity in the translation. For more details, check arXiv, identifier 2608.06240, released as a cross-type submission.
Key facts
- PRISM is a GAN-free flow-matching framework for unpaired image-to-image translation.
- It replaces global control with a learned per-feature gate.
- The gate's spatial prior uses standardized distance to target feature distribution.
- The gate controls both initialization and transport timing during ODE integration.
- The corruption is content-anchored (AdaIN) for structure-preserving translation.
- The paper is available on arXiv with ID 2608.06240.
- The announcement type is cross.
- The framework aims to separate content to keep from appearance to change.
Entities
Institutions
- arXiv