[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-84330-en":3,"doc-seo-84330-105":29,"detail-sidebar-cat-0-en-105":95},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},84330,687197207919,"Theodora","https://ap-avatar.wpscdn.com/avatar/a000253d6f5f7c60be?x-image-process=image/resize,m_fixed,w_180,h_180&k=1779446848396160552",8,"Research & Report","LUMI: Tokenizer-Agnostic LLM-Based Lossless Image Compression","Large language model (LLM)–based lossless image compression often encodes pixel values as text tokens processed by a pretrained model’s vocabulary head, which links coding performance to tokenizer behavior, numeric token conventions, and model-specific adaptation. LUMI (LLM-based Unified Model-agnostic lossless Image compression) introduces a tokenizer-agnostic framework using frozen LLM backbones: pixel embedding, intra-patch position encoding, soft-prefix parameters, and a 256-way prediction head are trained while the backbone remains fixed.","LUMI: Tokenizer-Agnostic LLM-Based Lossless Image Compression  \nChris Xing TIAN, Chengkai WU, Ziyu WANG, Rongqun LIN, Kecheng CHEN, Xiandong MENG, Haoliang LI, Shiqi WANG, and Siwei MA, Fellow, IEEE  \narXiv :2607 .0822 1v 1 [ cs .CV] 9 Jul 2026  \nAbstract—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, vocabularyspecific 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.  \nIndex Terms—Large language models, lossless image compression, entropy modeling, frozen foundation models.  \nI. INTRODUCTION  \nLarge language models (LLMs) are increasingly used as general sequence models beyond conventional text generation. In addition to natural language tasks, recent work has studied their use in coding [1], reasoning [2], retrieval [3], multimodal understanding [4], and data compression [5] . From the perspective of lossless compression [6], [7], this direction is motivated by a direct connection between autoregressive likelihood modeling and entropy coding, since a model that assigns accurate conditional probabilities to the next source symbol can be used by an arithmetic coder to obtain short codelengths.  \nChris Xing TIAN, Rongqun LIN, and Xiandong MENG are with Peng Cheng Laboratory, Shenzhen, China.  \nChengkai WU, Kecheng CHEN, and Haoliang LI are with the Department of Electrical Engineering, City University of Hong Kong, Hong Kong SAR.  \nZiyu WANG and Shiqi WANG are with the Department of Computer Science, City University of Hong Kong, Hong Kong SAR.  \nSiwei MA is with the School of Computer Science, Peking University, Beijing, China.  \nCorresponding author: Siwei MA (e-mail: [swma@pku.edu.cn](swma@pku.edu.cn)).  \nThis connection has motivated recent studies on LLMbased lossless image compression. Early approaches condition language models with visual prompts or residual information [8], [9], while pixel-level methods model RGB images as autoregressive sequences and use an LLM to predict the next pixel value. One example is P2-LLM, which represents pixel intensities as textual numeric tokens, provides prompt-based pixel priors, and adapts the LLM with parameter-efficient tuning for next-pixel prediction [10] . These studies indicate that pretrained LLMs can be used as entropy models for image compression.  \nHowever, the native interface of an LLM is not designed for exact image-symbol coding. A text tokenizer maps strings to vocabulary tokens, whereas a lossless RGB image codec needs probability distributions over the 25","cbCaiqSEyaZ3fzi7","https://ap.wps.com/l/cbCaiqSEyaZ3fzi7","pdf",1374899,1,10,"English","en",105,"# Introduction\n## Motivation for LLM-based lossless compression\n## Tokenizer dependency issues\n## Proposed tokenizer-agnostic formulation\n## Contributions overview","[{\"question\":\"为什么现有基于LLM的无损图像压缩会受到tokenizer影响？\",\"answer\":\"像素被转成文本数字 token 后，token 划分方式依赖具体 LLM 的 tokenizer，导致建模序列长度与算术编码所用的概率事件集合发生变化。此外，词表 logits 与像素字母表并不天然对齐。\"},{\"question\":\"LUMI如何实现与tokenizer无关的无损图像压缩？\",\"answer\":\"LUMI不再将像素当作文本 token，而是把原始像素强度、通道身份与补丁内二维位置映射到冻结 LLM 的连续嵌入空间，随后用冻结的解码器骨干处理，并通过 256 路预测头对下一像素的256种取值建模。\"},{\"question\":\"LUMI训练了哪些部分、冻结了哪些部分？\",\"answer\":\"只训练像素嵌入模块、补丁内位置编码、soft-prefix 参数以及预测头；LLM backbone 保持冻结不更新。\"},{\"question\":\"LUMI在不同LLM家族上的适用性如何体现？\",\"answer\":\"因为压缩接口直接接入 LLM 的嵌入与预测方式，绕过了对数值 token 分割的依赖，也不需要修改 LLM 内部权重，因此可在 LLaMA、Qwen、Gemma 等不同 tokenizer 家族上提供统一接口并提升跨域鲁棒性。\"}]",1784194864,25,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":90,"head_meta":92,"extra_data":94,"updated_unix":27},"lumi-tokenizer-agnostic-llm-based-lossless-image-compression","",{"@graph":35,"@context":89},[36,53,68],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/lumi-tokenizer-agnostic-llm-based-lossless-image-compression/84330/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-17","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81,85],{"name":72,"@type":73,"acceptedAnswer":74},"为什么现有基于LLM的无损图像压缩会受到tokenizer影响？","Question",{"text":75,"@type":76},"像素被转成文本数字 token 后，token 划分方式依赖具体 LLM 的 tokenizer，导致建模序列长度与算术编码所用的概率事件集合发生变化。此外，词表 logits 与像素字母表并不天然对齐。","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"LUMI如何实现与tokenizer无关的无损图像压缩？",{"text":80,"@type":76},"LUMI不再将像素当作文本 token，而是把原始像素强度、通道身份与补丁内二维位置映射到冻结 LLM 的连续嵌入空间，随后用冻结的解码器骨干处理，并通过 256 路预测头对下一像素的256种取值建模。",{"name":82,"@type":73,"acceptedAnswer":83},"LUMI训练了哪些部分、冻结了哪些部分？",{"text":84,"@type":76},"只训练像素嵌入模块、补丁内位置编码、soft-prefix 参数以及预测头；LLM backbone 保持冻结不更新。",{"name":86,"@type":73,"acceptedAnswer":87},"LUMI在不同LLM家族上的适用性如何体现？",{"text":88,"@type":76},"因为压缩接口直接接入 LLM 的嵌入与预测方式，绕过了对数值 token 分割的依赖，也不需要修改 LLM 内部权重，因此可在 LLaMA、Qwen、Gemma 等不同 tokenizer 家族上提供统一接口并提升跨域鲁棒性。","https://schema.org",{"og:url":51,"og:type":91,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":93,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":96},[97,101,105,109,114,119,124,127,132,135,138],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":98,"show_sort_weight":99,"slug":100},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":102,"show_sort_weight":103,"slug":104},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":106,"show_sort_weight":107,"slug":108},"Exam",70,"exam",{"id":110,"doc_module":4,"doc_module_name":45,"category_name":111,"show_sort_weight":112,"slug":113},5,"Comic",60,"comic",{"id":115,"doc_module":4,"doc_module_name":45,"category_name":116,"show_sort_weight":117,"slug":118},6,"Technology",50,"technology",{"id":120,"doc_module":4,"doc_module_name":45,"category_name":121,"show_sort_weight":122,"slug":123},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":125,"slug":126},30,"research-report",{"id":128,"doc_module":4,"doc_module_name":45,"category_name":129,"show_sort_weight":130,"slug":131},9,"Religion & Spirituality",20,"religion-spirituality",{"id":130,"doc_module":4,"doc_module_name":45,"category_name":133,"show_sort_weight":130,"slug":134},"World Cup","world-cup",{"id":21,"doc_module":4,"doc_module_name":45,"category_name":136,"show_sort_weight":21,"slug":137},"Lifestyle","lifestyle",{"id":139,"doc_module":4,"doc_module_name":45,"category_name":140,"show_sort_weight":110,"slug":141},19,"General","general"]