Liquid: Language Models are Scalable Multi-modal Generators

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

Introduction

Current Multimodal Large Language Models (MLLMs) often resemble "Frankenstein" architectures, stitching together disparate components like CLIP for vision understanding and Diffusion models for image generation. This modularity introduces complexity and prevents true cross-modal synergy. Liquid introduces a unified auto-regressive paradigm that enables standard LLMs to both "see" and "draw" by treating visual data as discrete tokens within an expanded vocabulary.

A Unified Multi-modal Paradigm

Unlike previous approaches that rely on externalized modules or separate heads for different tasks, Liquid achieves full unification. By continuing the training of pre-trained LLMs (such as Gemma-2 or Qwen2.5) with an extended vocabulary of 8,192 visual tokens, the researchers transformed these models into native multimodal generators.

Comparison of MLLM architectures. Liquid (f) achieves full unification across models, heads, representations, and objectives.

The technical implementation relies on a streamlined pipeline:

  1. Image Tokenization: Images are converted into 1,024 discrete codes using a VQGAN.
  2. Unified Token Space: Both text (BPE-tokenized) and image tokens reside in the same space.
  3. Next-Token Prediction: The model is trained with a single objective—predicting the next token, whether it is linguistic or visual.
The Liquid architecture: Text and images are converted into discrete tokens and processed within a unified token space by a standard LLM.

Scaling Laws and Cross-Modal Synergy

A key discovery of the Liquid project is the "Vanishing Multimodal Tax." In smaller models, adding visual tasks often degrades language performance. However, as model capacity scales (e.g., toward 32B parameters), this interference disappears. Furthermore, the researchers found that visual understanding and generation tasks are mutually beneficial; improving one task leads to gains in the other, suggesting a shared semantic foundation.

Key Results

Liquid demonstrates state-of-the-art performance across both generation and understanding benchmarks:

  • Image Generation: Liquid-7B achieved a Fréchet Inception Distance (FID) of 5.47 on the MJHQ-30K benchmark, surpassing specialized models like Stable Diffusion XL.
  • Multimodal Understanding: The model outperformed Chameleon on understanding tasks and matched the performance of established models like LLaVA.
  • Efficiency: By leveraging "continue training" on existing LLMs rather than training from scratch, Liquid is approximately 100x more cost-effective than previous unified attempts.

Conclusion

Liquid proves that pre-trained LLMs are inherently capable of becoming high-quality multimodal generators. By unifying the architecture and training objective, Liquid paves the way for more efficient, scalable, and synergistic multimodal AI systems. Future work will focus on refining tokenization to better align visual and linguistic semantics and scaling to higher resolutions.

Summary updated

BibTeX
@article{wu2026liquid,
  author = {Junfeng Wu and Yi Jiang and Chuofan Ma and Yuliang Liu and Hengshuang Zhao and Zehuan Yuan and Song Bai and Xiang Bai},
  title = {{Liquid: Language Models are Scalable Multi-modal Generators}},
  journal = {Int. J. Comput. Vis.},
  year = {2026},
  doi = {10.1007/s11263-025-02639-5},
}
Figure