LongCat-Image Technical Report

TL;DR

LongCat-Image is Meituan’s open-source bilingual (Chinese-English) image-generation foundation model — 6B parameters, Apache 2.0, with a companion LongCat-Image-Edit variant for instruction-guided editing. The release explicitly targets three gaps in current open T2I: multilingual text rendering (especially complex/rare Chinese characters), photorealism, and deployment efficiency at sub-10B scale. Method centers on rigorous multi-stage data curation (pre-training / mid-training / SFT) plus an RL phase with curated reward models. Positioned by the authors as a new SOTA on Chinese character rendering with industry-leading dictionary coverage, evaluated on GenEval, DPG, WISE, GlyphDraw2, CVTG-2K, ChineseWord, ImgEdit, GEdit-EN, and GEdit-CN.

Key claims

The model is a 6B-parameter bilingual (Chinese-English) image generation foundation model, released open-source under Apache 2.0 with code, weights, and a separate Image-Edit variant [§1, Abstract].
Despite the 6B size, the authors claim it surpasses open-source models several times larger across multiple benchmarks (GenEval, DPG, WISE for general T2I; GlyphDraw2, CVTG-2K, ChineseWord for text rendering; ImgEdit, GEdit-EN, GEdit-CN for editing) [Figure 1].
The model achieves industry-leading coverage of the Chinese dictionary and superior accuracy on rare/complex Chinese characters versus existing SOTA open and commercial solutions [Abstract, §2.1.2].
Text rendering relies on a specialized character-level encoding strategy triggered by enclosing target text in single/double quotation marks (English '...'/"..." or Chinese ‘…’/“…”) — without explicit quotation marks the mechanism does not activate and rendering quality degrades severely [HF model card §Special Handling for Text Rendering].
Data pipeline is multi-stage: pre-training, mid-training, SFT, plus a coordinated RL phase using curated reward models — the explicit position is that quality comes from data curation discipline rather than parameter count [Abstract, §2].
Data curation modules cover filtering, meta-information extraction, multi-granularity captioning, and stratification, with a separate data synthesis stage [§2.1, §2.2].
Inference recipe ships with enable_cfg_renorm=True and enable_prompt_rewrite=True by default in the diffusers pipeline, and runs in ~17 GB VRAM with model-CPU-offload — a deliberate consumer-hardware target [HF model card §Run Text-to-Image Generation].

Method

The release is integrated into HuggingFace diffusers as LongCatImagePipeline, with default inference at 50 steps, guidance scale 4.0, and configurable resolution (the model card shows 768×1344). Two operational defaults are notable: CFG renormalization is on by default, and a prompt-rewrite step is enabled in the pipeline — i.e. the user prompt is rewritten by a stronger LLM before the diffusion backbone consumes it, the same recipe pattern as Z-Image’s Prompt Enhancer. The character-level encoding for quoted text is a model-specific mechanism: the tokenizer treats content inside matched quotation marks as a separate character-level stream so that downstream attention can attend to individual glyphs — without quotes, the model falls back to ordinary BPE text tokens, which is insufficient for rare Chinese characters.

The training pipeline (per the tech report’s TOC) is structured as four data-construction stages (filtering, meta-info extraction, multi-granularity captioning, stratification) plus data synthesis, followed by pre-training, mid-training, SFT, and an RL phase with curated reward models. The full architectural and quantitative detail (model architecture diagrams, exact data scale, RL formulation, headline benchmark numbers) lives in the arXiv PDF (2512.07584); the HF model card and abstract are this page’s primary fetched sources.

Results

The abstract and HF card frame the headline result as new-SOTA Chinese character rendering — both in coverage (supporting “even complex and rare characters”) and accuracy — versus open-source models and commercial systems on benchmarks including GlyphDraw2, CVTG-2K, and ChineseWord [Abstract, Figure 1]. General T2I quality is evaluated on GenEval, DPG, and WISE; editing on ImgEdit, GEdit-EN, and GEdit-CN [Figure 1]. The release additionally claims photorealism gains attributable to the data-curation + RL-with-curated-reward-models pipeline rather than to model scale. Exact benchmark numbers are not included in the abstract or model card and would require reading the PDF [§5 not in fetched body].

Why it’s interesting

This is the third datapoint in a short-term trend toward open ~6B bilingual image foundation models with Chinese-text-rendering as a headline capability — alongside Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer (Alibaba Tongyi-MAI, 6B, single-stream DiT, also bilingual) and Qwen-Image Technical Report (Qwen, 20B, MMDiT). LongCat-Image and Z-Image are direct competitors at the same parameter count and same bilingual EN/ZH target; the architectural contrast (LongCat-Image’s reported FLUX-like architecture per the poster’s note vs Z-Image’s claimed Single-Stream S3-DiT) and the data-curation-vs-architecture framing are worth tracking as the second-half-2025 “small efficient open T2I” cohort consolidates. This also continues the Open foundation-model releases pattern of Chinese-lab releases that explicitly position against 20-80B open competitors on quality-per-parameter rather than absolute quality — the same axis Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer used as its counterargument to HunyuanImage 3.0 Technical Report.