Reproduce “Learning Anisotropic Prediction Targets for Flow Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Rayhan Zirvi, Hansheng Chen, Gordon Wetzstein, Leonidas Guibas
AnisoFlow learns direction-specific noise weights in a learned basis, changing a flow model’s prediction target without changing its architecture. It converts the network’s output back into velocity. On ImageNet 256×256, JiT-H/16’s FID falls from 1.90 to 1.55. Gains are strongest in pixel space; tested latent models remain close to default targets, and measured inference-step overhead is 0.6%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.