LUMI: Tokenizer-Agnostic LLM-Based Lossless Image Compression

2026-07-09

AI Summary

This work addresses the limitations of existing large language model (LLM)-based lossless image compression methods, which rely on text-specific tokenizers and consequently suffer from poor model adaptability and weak cross-domain robustness. The authors propose LUMI, a novel framework that achieves tokenizer-free LLM-based image compression for the first time. LUMI maps raw pixels into the continuous embedding space of a frozen LLM (e.g., LLaMA, Qwen, or Gemma) via a pixel embedding module, incorporates intra-block 2D positional encoding, employs a native 256-way pixel prediction head, and leverages soft prefix tuning. By reframing the task from language symbol modeling to pixel-space adaptation, LUMI achieves state-of-the-art compression ratios across natural, medical, and remote sensing images, significantly outperforming tokenizer-dependent baselines and demonstrating strong generalization across both models and domains.

Abstract

Large language model (LLM)-based lossless image compression methods typically represent pixel data through the native text interface of a pretrained model, converting pixel values into token sequences that the LLM processes through its vocabulary head. This design shows that pretrained language models can provide probability estimates for image coding, but it also couples compression to tokenizer behavior, vocabulary-specific numeric tokens, and model-family-specific adaptation. In this paper, we present LUMI (LLM-based Unified Model-agnostic lossless Image compression), a tokenizer-agnostic framework for lossless RGB image compression with frozen LLM backbones. LUMI replaces pixel-as-text tokenization with a pixel embedding module that maps raw intensity and channel information into the continuous embedding space of the LLM. It further introduces intra-patch position encoding to retain two-dimensional spatial structure after flattening, and uses a 256-way prediction head to produce probabilities over the native pixel alphabet. Only the pixel embedding, position encoding, soft-prefix parameters, and prediction head are trained, while the LLM backbone remains fixed. Experiments on natural, medical, and remote-sensing image benchmarks with LLaMA, Qwen, and Gemma backbones show that LUMI provides a unified interface across tokenizer families, achieves competitive compression rates, and improves cross-domain robustness over tokenizer-based LLM compression baselines. These results formulate LLM-based lossless image compression as pixel-space adaptation of frozen foundation models rather than tokenizer-specific language-symbol modeling.

Problem

• lossless image compression
• large language models
• tokenizer dependency
• pixel representation
• model agnosticism

Innovation

• tokenizer-agnostic
• LLM-based compression
• pixel embedding
• lossless image compression
• frozen foundation model