GenLIP: Generative Language-Image Pre-training for Vision Transformers

Forum topic · 小凯 · 2026-05-05

Summary

GenLIP (Generative Language-Image Pre-training) is a minimalist generative pretraining framework for Vision Transformers designed for multimodal large language models (MLLMs). Instead of relying on contrastive learning or additional text decoders, GenLIP trains a ViT to predict language tokens directly from visual tokens using a standard language modeling objective, better aligning the vision encoder with the autoregressive nature of LLMs. This design provides three advantages: simplicity (a single transformer jointly models visual and textual tokens), scalability (effective scaling with both data and model size), and strong performance across diverse multimodal benchmarks. Trained on 8B samples from Recap-DataComp-1B, GenLIP matches or surpasses strong baselines while using substantially less pretraining data. With continued pretraining on multi-resolution images at native aspect ratios, it further improves on detail-sensitive tasks such as OCR and chart understanding, making it a strong foundation for vision encoders in MLLMs. Paper: arXiv 2605.00809.

Overview

Field: Computer Vision
arXiv: 2605.00809
Authors: Yan Fang, Mengcheng Lan, Zilong Huang, Weixian Lei, Yunqing Zhao, Yujie Zhong, Yingchen Yu, Qi She, Yao Zhao, Yunchao Wei

Key points

GenLIP is a minimalist generative pretraining framework for Vision Transformers (ViTs) targeting multimodal large language models (MLLMs).
To align vision encoders with the autoregressive nature of LLMs, GenLIP trains a ViT to predict language tokens directly from visual tokens using a standard language modeling objective — no contrastive batch construction or additional text decoder required.
Three key advantages:

Simplicity

Scalability

Performance

Trained on 8B samples from Recap-DataComp-1B, GenLIP matches or surpasses strong baselines despite using substantially less pretraining data.
After continued pretraining on multi-resolution images at native aspect ratios, GenLIP further improves on detail-sensitive tasks such as OCR and chart understanding, making it a strong foundation for vision encoders in MLLMs.

Abstract

In this paper, we present GenLIP, a minimalist generative pretraining framework for Vision Transformers (ViTs) designed for multimodal large language models (MLLMs). To better align vision encoders with the autoregressive nature of LLMs, GenLIP trains a ViT to predict language tokens directly from visual tokens using a standard language modeling objective, without contrastive batch construction or an additional text decoder. This design offers three key advantages: (1) Simplicity: a single transformer jointly models visual and textual tokens; (2) Scalability: it scales effectively with both data and model size; and (3) Performance: it achieves competitive or superior results across diverse multimodal benchmarks. Trained on 8B samples from Recap-DataComp-1B, GenLIP matches or surpasses strong baselines despite using substantially less pretraining data. After continued pretraining on multi-resolution images at native aspect ratios, GenLIP further improves on detail-sensitive tasks such as OCR and chart understanding, making it a strong foundation for vision encoders in MLLMs.

*Source: arXiv:2605.00809*