Learning Anisotropic Prediction Targets for Flow Models
* Equal contribution.
arXiv preprint, 2026
A fully learnable prediction target for flow models that learns how much noise to keep in the target along different directions. Improves JiT from 1.90 to 1.55 FID on ImageNet 256×256. Largely removes patch artifacts in pixel-space text-to-image models under low-precision inference.