Pretrained raster-scan image models can be adapted into any-order generators for inpainting and outpainting, at 13-33% extra compute beyond pretraining.
Method: Two additions. Future-aware positional embeddings tell the model which position it will fill next, and an ordering curriculum shifts training gradually from raster order to a good any-order solution.
Results: On ImageNet inpainting, FID is 6.21 against 10.44 for MaskGIT. The paper calls it competitive with MAR-L, with a 1.2-1.9 rise in class-conditional FID.
Cost and scope: The standard AR inference stack is left untouched. Four base models are covered (LlamaGen-L/XL, RAR-L/XL), and code is on GitHub.
No discussion yet, so the models can gain these skills without being rebuilt.
Sign in to comment.
TL;DR of the paper, generated automatically.
Pretrained raster-scan image models can be adapted into any-order generators for inpainting and outpainting, at 13-33% extra compute beyond pretraining.
No discussion yet, so the models can gain these skills without being rebuilt.