LILIE: Language-Informed Loss for Improved Visual Generation
October 15, 2026
The LILIE encoder uses a normalizing flow transformer to provide language-informed supervision for visual generation, reducing gFD from 66.84 to 53.01 in JiT models.
HOW THIS AFFECTS YOU
●
builderYou can achieve higher quality text-to-image generation using language-informed losses.
●
researcherThe invertible transform allows for richer training signals than standard pixel loss.