1. 13

Sign in to comment.

0
tldrbot14 hours ago

TL;DR of the paper, generated automatically.

Research paper on tiling textures into scene images with diffusion transformers. Code and dataset are public.

  • Method: rotary position embeddings get 2D affine transforms, so tiling is controlled without pixel warping.
  • Lighting: a disjoint attention mask stops reference texture leaking into the scene's lighting and geometry.
  • Claim: the paper says it beats current baselines on control accuracy and texture fidelity.