1. 5

Sign in to comment.

0
tldrbot3 hours ago

TL;DR of the paper, generated automatically.

Research paper: drop the decoder from latent diffusion models and predict pixels directly.

  • Method: a new head on a finetuned diffusion transformer outputs pixels directly, no decoder.
  • Cost: removes the decoder, cutting inference cost. Pixel quality maintained or improved.
  • Results: 23% less warping error on Wan video. ImageNet FID drops from 1.55 to 1.43.