Training an image generator involves teaching a network to predict a specific quantity, like noise, clean data, or a velocity, from a noisy image. The choice of target matters a lot, but it has always been made by hand. This paper introduces AnisoFlow, which learns the best target automatically, assigning a different weight to each direction in the data. On ImageNet 256x256, AnisoFlow improved the Frechet Inception Distance (FID) from 1.90 to 1.55, showing a meaningful gain in image quality without changing the network architecture.
Image generators improve without architecture changes by learning what to predict
Zirvi and colleagues report better pixel-based image generation by learning prediction targets, while leaving the network architecture, training loss and sampling procedure unchanged.
Top University
Rayhan Zirvi · Hansheng Chen · Gordon Wetzstein · Leonidas Guibas
Stanford University · California Institute of Technology
Research Digest··2 min read

In one line
Replacing a fixed prediction target with a learned anisotropic one improves flow model image quality on ImageNet and text-to-image tasks.
Before you read this
An idea this paper builds onRead it as
Image generators are commonly trained to predict noise, a clean image or the direction connecting them, and that choice affects image quality.
The authors developed Anisotropic Flow Models, or AnisoFlow, which learns how much noise to include in its training target along different directions in the image representation.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.