End-to-end visual generation with pixel-space loss has recently achieved competitive performance without the need for learned image tokenizers. In this work, we study whether language-informed supervision can provide a richer training signal than pixel loss for visual generation. We introduce LILIE (pronounced “lily”), a Language-Informed and Lossless Image Encoder. LILIE is an invertible image transform trained such that a vision-language model can recover captions from noised features. The encoder uses a normalizing flow transformer trained jointly with a vision-language model via next token prediction. We then use the encoder to extract a simple language-informed loss for text-to-image generation. We train JiT models with LILIE loss, reducing gFD by 14 points (66.84 → 53.01) against a pixel-only baseline on the GPIC dataset.
LILIE is an invertible encoder built with the TARFlow architecture. We train the encoder by passing an image through the encoder, corrupting the encoded image with additive Gaussian noise, then captioning the noised features with a LoRA-finetuned VLM. The objective is a standard cross-entropy loss against ground-truth captions from the GPIC corpus.
Once we have LILIE, we use L2 distance in LILIE space to supervise JiT training. LILIE loss works on its own and improves generation quality over a pixel-only baseline (66.84 → 55.78 gFD), and combining LILIE loss with pixel loss further improves gFD to 53.01.
Click a pair to zoom in.
































@misc{agarwal2026lilie,
title={A Language-Informed Loss for Visual Generation},
author={Suchir Agarwal and Keshigeyan Chandrasegaran and Ziyu Chen and Kyle Sargent and Juan Carlos Niebles and Li Fei-Fei and Jiajun Wu},
year={2026},
eprint={2610.23653},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2610.23653},
}