Stanford's LILIE: language-informed lossless image encoder cuts T2I gFD by 14 points
jcniebles · x · 2026-10-20
Stanford researchers introduce LILIE, a Language-Informed and Lossless Image Encoder trained with language supervision to replace pixel-space loss in visual generative models. Pixel loss treats all pixels equally even though they contribute unevenly to semantics; LILIE is an invertible normalizing-flow transform trained jointly with a VLM via caption recovery from noised features. Using LILIE-space loss to train JiT models improves gFD on GPIC from 66.84 to 53.01 (14 points), and combining the losses improves further. Authors include Suchir Agarwal, Fei-Fei Li, and Jiajun Wu.
More from Multimodal
- Building 3D worlds with Gemini 4 Argon and Blender is my new favourite hobby — jacalulu · 2026-10-20
- Zelda meets Pikachu in a slick AI-generated video powered by Breeze TTS2 and Krea 2 — dramaton42 · 2026-10-20
- AI-generated boxing sparring video fools viewers with realism — Electrical-Theme2847 · 2026-10-20
- One-shot ASCII poster generated in a single Opus prompt — zsakib_ · 2026-10-20
- Ant Group's Ming-3.0-tts tops open-weight TTS leaderboards with 93.2% pronunciation robustness — ArtificialAnlys · 2026-10-20
- 'I'm tired boss': a hyper-realistic AI-generated soldier photo fools the timeline — flowersslop · 2026-10-20