Scholarlink

<p>We present Liquid, a versatile and native auto-regressive generation paradigm that seamlessly integrates visual comprehension and generation by tokenizing images into discrete codes and learning these code embeddings alongside text tokens within a shared feature space for both vision and language. Unlike previous multimodal large language model (MLLM), Liquid achieves this integration using any existing large language models (LLMs), eliminating the need for external pretrained visual modules such as CLIP and diffusion models. For the first time, Liquid reveals that the power-law scaling laws of unified multimodal models align with those observed in language models, and it discovers that the trade-offs between visual and language tasks diminish as model size increases. Furthermore, the unified token space enables visual generation and comprehension tasks to mutually enhance each other, effectively removing the typical interference seen in earlier models. We demonstrate that existing LLMs can serve as strong foundations for Liquid, saving training costs by 100<i>times</i> while surpassing Chameleon in multimodal capabilities. Compared to previous unified multimodal models, Liquid maintains on-par language performance comparable to mainstream LLMs like Llama2, preserving its potential as a foundational model. Building on this foundation, Liquid outperforms visual generation models like SD v2.1 and SD-XL (FID of 5.47 on MJHQ-30K), excelling in both vision-language and text-only tasks. The code and models are available at <a href="https://github.com/FoundationVision/Liquid">https://github.com/FoundationVision/Liquid</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Liquid: Language Models are Scalable and Unified Multi-Modal Generators

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Junfeng Wu,
  • Yi Jiang,
  • Chuofan Ma,
  • Yuliang Liu,
  • Hengshuang Zhao,
  • Zehuan Yuan,
  • Song Bai,
  • Xiang Bai

摘要

We present Liquid, a versatile and native auto-regressive generation paradigm that seamlessly integrates visual comprehension and generation by tokenizing images into discrete codes and learning these code embeddings alongside text tokens within a shared feature space for both vision and language. Unlike previous multimodal large language model (MLLM), Liquid achieves this integration using any existing large language models (LLMs), eliminating the need for external pretrained visual modules such as CLIP and diffusion models. For the first time, Liquid reveals that the power-law scaling laws of unified multimodal models align with those observed in language models, and it discovers that the trade-offs between visual and language tasks diminish as model size increases. Furthermore, the unified token space enables visual generation and comprehension tasks to mutually enhance each other, effectively removing the typical interference seen in earlier models. We demonstrate that existing LLMs can serve as strong foundations for Liquid, saving training costs by 100times while surpassing Chameleon in multimodal capabilities. Compared to previous unified multimodal models, Liquid maintains on-par language performance comparable to mainstream LLMs like Llama2, preserving its potential as a foundational model. Building on this foundation, Liquid outperforms visual generation models like SD v2.1 and SD-XL (FID of 5.47 on MJHQ-30K), excelling in both vision-language and text-only tasks. The code and models are available at https://github.com/FoundationVision/Liquid.