A Language-Informed Loss for Visual Generation
Summary
该论文提出 LILIE（Language-Informed and Lossless Image Encoder），利用可逆归一化流编码器为文本到图像生成提供语言监督损失，在 GPIC 数据集上将 JiT 模型的 gFD 从 66.84 降至 53.01，相比纯像素损失显著提升保真度。
View Cached Full Text
Cached at: 10/21/26, 07:04 AM
Paper page - A Language-Informed Loss for Visual Generation
Source: https://huggingface.co/papers/2610.23653
Abstract
LILIE uses an invertible normalizing-flow encoder to derive language-informed supervision for text-to-image generation, improving fidelity over pixel-only training.
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 aninvertible image transformtrained such that avision-language modelcan recover captions from noised features. The encoder uses anormalizing flowtransformer trained jointly with avision-language modelvianext token prediction. We then use the encoder to extract a simplelanguage-informed lossfortext-to-image generation. We train JiT models with LILIE loss, reducing gFD by 14 points (66.84 to 53.01) against a pixel-only baseline on the GPIC dataset. Our code, encoders, and generative models are available at https://suchiragarwal.com/lilie.
View arXiv pageView PDFProject pageAdd to collection
Get this paper in your agent:
hf papers read 2610\.23653
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper1
#### suchiragarwal/lilie Image Feature Extraction• Updatedabout 12 hours ago • 2
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2610.23653 in a dataset README.md to link it from this page.
Spaces citing this paper1
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
AFP-GIC: Controllable Generative Image Compression [R]
AFP-GIC, a controllable generative image compression framework published in IEEE Access (2026), uses an asymmetric Adaptive Fused Prior Transfer pipeline to reduce hallucinations and local distortion at ultra-low bitrates, achieving 18.1% lower decoder latency and 20.5% fewer parameters than DC-VIC, with released code, benchmark data, and an interactive Hugging Face demo.
Training Text-to-Image Models 3.6× Faster
Linum AI introduces JiT-DDT, a novel encoder-decoder architecture that trains text-to-image models 3.6× faster than previous methods while generating images with higher resolution.
Adaptive Fused Prior Transfer for Controllable Generative Image Compression
AFP-GIC is an IEEE Access 2026 paper proposing Adaptive Fused Prior Transfer for controllable generative image compression, achieving up to 167:1 compression with a single pretrained model across five bitrate operating points and reducing decoder latency by 18.1% versus DC-VIC.
GRACE: Generation-aware latent compression for efficient video generation
GRACE提出两阶段生成感知潜空间压缩框架，在保持与预训练DiT兼容的前提下将视频自编码器压缩率提升，使Wan2.1-I2V-14B的token数减少8倍、延迟降低11.1倍，同时在VBench上保持同等生成质量。
On the Design Fundamentals of Pixel Text Representation Learning
Pixel Linguist II improves visual text representation learning via variable resolution training, natural image-text grounding, layout-aware rendering, and multilingual curricula, achieving state-of-the-art results and robust compression performance.