A Language-Informed Loss for Visual Generation
LILIE cut gFD from 66.84 to 53.01 against a pixel-only baseline on GPIC.
The researchers built an invertible image encoder trained to help a vision-language model recover captions from noised image features. They used that encoder to derive a language-informed training loss for text-to-image generation. JiT models trained with this loss achieved the reported improvement. The authors say their code, encoders and generative models are available.
HF Daily Papers' note
The researchers built an invertible image encoder trained to help a vision-language model recover captions from noised image features. They used that encoder to derive a language-informed training loss for text-to-image generation. JiT models trained with this loss achieved the reported improvement. The authors say their code, encoders and generative models are available.
HF Daily Papers' note
score 5