arXiv:2611.17767v1 Announce Type: new Abstract: Flow-based generative models are trained on a hand-chosen prediction target, such as the velocity, the noise, or the clean data. This choice strongly affects training and sample quality. We introduce Anisotropic Flow Models (AnisoFlow), which replaces a fixed prediction target with a fully learnable one. AnisoFlow learns to scale the noise component of the target anisotropically, assigning a separate nonnegative weight to each direction in a jointly learned basis. A closed-form map recovers the velocity from the network's prediction, leaving the architecture, flow-matching loss, and sampler unchanged. Since recovery is exact, training optimizes the target purely to make it easier for the network to predict. We theoretically analyze the resulting prediction error and predict that the preferred noise weight in each direction scales near-linearly with data variance, in agreement with the weights learned empirically in pixel space. AnisoFlow improves FID from 1.90 to 1.55 over the JiT baseline on ImageNet 256$\times$256, by learning only the prediction target, without representation alignment or architectural changes. In pixel-space text-to-image generation, AnisoFlow improves image quality and, through better-conditioned velocity recovery, largely removes patch-seam artifacts under reduced-precision inference.
Learning Anisotropic Prediction Targets for Flow Models
AnisoFlow, a flow-based generative modeling method that replaces a fixed prediction target with a fully learnable, anisotropically scaled one, improved FID from 1.90 to 1.55 over the JiT baseline on ImageNet 256×256, according to the arXiv paper 2611.17767v1. The method assigns a separate nonnegative weight to each direction in a jointly learned basis and uses a closed-form map to recover velocity from the network's prediction, leaving the architecture, flow-matching loss, and sampler unchanged. The authors report that the preferred noise weight in each direction scales near-linearly with data variance, matching weights learned empirically in pixel space, and that in pixel-space text-to-image generation AnisoFlow largely removes patch-seam artifacts under reduced-precision inference.
Run your AI side-project on zahid.host
EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.