Stanford's LILIE loss cuts text-to-image gFD by 14 points over pixel-only training
StanfordUniversity · hf · 2026-10-21
A Stanford paper (authors include Fei-Fei Li and Jiajun Wu) introduces LILIE, a Language-Informed and Lossless Image Encoder that studies whether language-informed supervision gives richer training signals than pixel-space loss.
LILIE is an invertible normalizing-flow transformer trained so a vision-language model can recover captions from noised features via next-token prediction; the encoder then yields a language-informed loss for text-to-image training. Training JiT models with the LILIE loss reduces gFD from 66.84 to 53.01 (14 points) versus a pixel-only baseline on GPIC. Code, encoders, and models are released.
More from Multimodal
- Open-source YuE2 claims Suno-level music generation with symbolic planning — drdanielbender · 2026-10-21
- Gradium open-sources Phonon, sub-50M TTS models under Apache 2.0 — mishig25 · 2026-10-21
- AI still life lesson: a plastic dust sheet beats a list of mood words — tisch_eins · 2026-10-21
- Nano Banana 2.1 character renders surface with Midjourney-style dark fairytale looks — azed_ai · 2026-10-21
- Ex-Google Brain/DeepMind team launches Vyb, an AI design model with editable layered output — testingcatalog · 2026-10-21
- Making a classroom case-explainer video with Opus 5.5 — technollama · 2026-10-21