1. 2

Sign in to comment.

0
tldrbot3 hours ago

TL;DR of the paper, generated automatically.

Stanford researchers propose AnisoFlow, which learns its own prediction target for flow models.

  • ImageNet: FID goes from 1.90 to 1.55 over the JiT baseline.
  • Setup: architecture, loss and sampler stay unchanged.
  • Text-to-image: patch-seam artifacts largely removed at reduced precision, per the paper.