Megadose The AI curve, compressed.

Learning Anisotropic Prediction Targets for Flow Models

· ArXiv · AI/CL/LG ·
Learning only the prediction target cut ImageNet 256×256 FID from 1.90 to 1.55 against the JiT baseline.

AnisoFlow replaces a fixed target, such as velocity or noise, with a learned target that weights noise separately along directions in a learned basis. An exact conversion recovers velocity while leaving the architecture, flow-matching loss and sampler unchanged. The authors’ analysis predicts that each direction’s preferred noise weight scales roughly linearly with data variance, matching their pixel-space experiments. In pixel-space text-to-image tests, they also report improved image quality and largely eliminated patch-seam artifacts during reduced-precision inference.

ArXiv · AI/CL/LG's note

score 5

Categories: Research