Skip to main content
Back

Stanford's AnisoFlow Boosts Image Generation FID From 1.90 to 1.55

AnisoFlow makes the prediction target of flow models fully learnable, cutting ImageNet FID from 1.90 to 1.55 and killing patch seams in 8-bit inference.

  • AnisoFlow makes the flow-model prediction target fully learnable via a per-direction noise-weighting operator.
  • Improves JiT-H/16 on ImageNet 256 from 1.90 to 1.55 FID with no architecture changes.
  • Closed-form velocity recovery keeps the flow-matching loss and sampler unchanged during training.
  • Learned weights empirically scale near-linearly with data variance, matching the paper's theoretical prediction.
  • Eliminates patch-seam artifacts under INT8 and FP8 inference by conditioning on recovered velocity error.
  • Code promised on LakonLab; project page at hanshengchen.com/anisoflow.

Flow-based image generators can train a network to predict clean data, noise, or the velocity connecting them. Researchers at Stanford and Caltech report in the AnisoFlow paper that learning this target improves ImageNet 256 × 256 FID from 1.90 to 1.55 after 600 epochs on the same JiT-H/16 backbone. Lower FID indicates a closer match between generated and reference image distributions.

The proposed method, AnisoFlow, keeps the backbone, interpolation path, flow-matching objective, and sampling solver fixed. Training adds a learnable operator that defines the prediction target, plus a closed-form conversion from that target to the velocity required by the loss and sampler.

The target becomes a learned preconditioner

AnisoFlow defines the target as uW = Wε - x, where x is clean data, ε is noise, and W is a learned positive semidefinite operator. The operator can be written as W = Q diag(wi) QT, with a learned basis Q and non-negative directional weights wi. “Anisotropic” means that each direction can receive a different noise weight.

A closed-form recovery map converts the model’s output and current noisy state into the standard velocity. The flow-matching loss evaluates that recovered velocity, and the existing sampler consumes it during inference. The learned weights and basis are optimized jointly with the network.

Target geometry changes the regression problem

Flow matching constructs a path such as xt = (1 - t)x + tε and trains a vector field to move samples along it. Target parameterization changes the scale, variance, and numerical conditioning of the regression problem, including how prediction errors are amplified when converted back to velocity.