PaperlayerView exact revision on arXiv

Source-linked AI summary

LongCat-Image Technical Report

Meituan LongCat Team, Hanghang Ma, Haoxian Tan, Jiale Huang, Junqiang Wu, Jun-Yan He, Lishuai Gao, Songlin Xiao, Xiaoming Wei, Xiaoqi Ma, Xunliang Cai, Yayong Guan, Jie Hu

arXiv:2512.07584v1cs.CV

TL;DR

LongCat-Image addresses persistent gaps in image generation and editing, including photorealism, multilingual text rendering, efficiency, and visual consistency. It combines a lightweight 6B diffusion architecture with curated multi-stage training and reports strong generation, Chinese rendering, and editing results while releasing models, checkpoints, and training code.

  • Problem

    Existing systems face unresolved challenges in precise image editing, visual consistency, costly parameter scaling, and limited caption information density.

  • Method

    LongCat-Image combines a 6B diffusion architecture with systematic data curation, staged alignment, specialized Chinese-rendering strategies, and task-specific editing training.

  • Results

    LongCat-Image reports strong photorealism, industry-leading Chinese character coverage and accuracy, and state-of-the-art open-source image-editing performance with visual consistency.

  • Takeaways & Limitations

    The released models, intermediate checkpoints, and complete training code provide an open development ecosystem for image generation and editing research.

  • Takeaways & Limitations

    Long-tail editing data remains limited because mining valid samples from web-scale interleaved corpora is extremely resource-intensive.

Abstract

from arXiv · show

We introduce LongCat-Image, a pioneering open-source and bilingual (Chinese-English) foundation model for image generation, designed to address core challenges in multilingual text rendering, photorealism, deployment efficiency, and developer accessibility prevalent in current leading models. 1) We achieve this through rigorous data curation strategies across the pre-training, mid-training, and SFT stages, complemented by the coordinated use of curated reward models during the RL phase. This strategy establishes the model as a new state-of-the-art (SOTA), delivering superior text-rendering capabilities and remarkable photorealism, and significantly enhancing aesthetic quality. 2) Notably, it sets a new industry standard for Chinese character rendering. By supporting even complex and rare characters, it outperforms both major open-source and commercial solutions in coverage, while also achieving superior accuracy. 3) The model achieves remarkable efficiency through its compact design. With a core diffusion model of only 6B parameters, it is significantly smaller than the nearly 20B or larger Mixture-of-Experts (MoE) architectures common in the field. This ensures minimal VRAM usage and rapid inference, significantly reducing deployment costs. Beyond generation, LongCat-Image also excels in image editing, achieving SOTA results on standard benchmarks with superior editing consistency compared to other open-source works. 4) To fully empower the community, we have established the most comprehensive open-source ecosystem to date. We are releasing not only multiple model versions for text-to-image and image editing, including checkpoints after mid-training and post-training stages, but also the entire toolchain of training procedure. We believe that the openness of LongCat-Image will provide robust support for developers and researchers, pushing the frontiers of visual content creation.

1 Introduction

LongCat-Image targets the quality, efficiency, text-rendering, and editing limitations of current image-generation systems with a lightweight 6B diffusion model and staged training strategy. It reports strong photorealism, Chinese text rendering, editing consistency, and community accessibility.

  • Current image-editing systems still struggle with precise instruction execution and strict visual consistency between original and edited images.
  • Unrestrained parameter growth has increased computational costs and deployment barriers without delivering the anticipated qualitative breakthrough.
  • LongCat-Image uses a 6B-parameter lightweight diffusion architecture to balance generation capability with training and inference efficiency.The model combines a hybrid MM-DiT and Single-DiT structure with Qwen2.5VL-7B as its text encoder.
  • Systematic data curation and staged training support remarkable photorealism, while dedicated data, architecture, and training strategies improve Chinese character rendering.
  • The editing model achieves state-of-the-art open-source performance in instruction following, image quality, and visual consistency.
  • The project releases final models, intermediate checkpoints, and complete training code to lower barriers for research and development.

2 Data

LongCat-Image builds a 1.2-billion-sample corpus through filtering, metadata extraction, multi-granularity captioning, aesthetic assessment, and stratification. The data design addresses caption limitations by combining structured metadata with hierarchical and photographic descriptions.

  • 1.2 billion samples form the curated training corpus, whose composition is summarized in the data overview.
  • The curation pipeline filters low-quality and duplicate samples, extracts metadata, recaptions images, and stratifies data for multi-stage training.
  • Deduplication combines MD5 hashing for exact duplicates with SigLIP similarity assessment for near-duplicate entries.
  • Filtering: Filtering excludes images with shortest edges below 384 pixels, aspect ratios outside 0.25–4.0, visible watermarks, and LAION-Aesthetics scores below 4.5.
  • Filtering: AIGC detection removes synthetic data because even a small fraction can disrupt optimization and produce plastic or greasy textures.
  • Meta Information Extraction: Metadata covers category, style, named entities, OCR text, and aesthetics to support balanced organization and more informative captions.
  • Aesthetic Evaluation: Aesthetic evaluation separates technical Quality from artistic Artistry and uses complementary assessment methodologies rather than a single metric.
  • Multi-Granularity Captioning: Existing VLM captioning is limited by weak world-knowledge integration, repetitive formats, and low information density from verbose descriptions.

3 Model Design

The model uses a FLUX.1-dev-inspired transformer with double- then single-stream attention, a FLUX.1-dev VAE, and Qwen2.5VL-7B conditioning. Character-level tokenization, 3D MRoPE, and built-in prompt refinement support multilingual generation and editing.

  • The transformer uses double-stream attention in early layers and single-stream attention later, with approximately a 1:2 block ratio.
  • The FLUX.1-dev VAE provides reconstruction fidelity for fine typography and intricate textures.
  • Input images undergo 8× spatial compression and 2 × 2 token merging before entering the DiT module.
  • Qwen2.5VL-7B embeds prompts into continuous representations that condition the DiT model for multilingual generation and editing.
  • Character-level tokenization is applied to quoted visual text to reduce generation complexity, improve data efficiency, and accelerate convergence.
  • A 3D MRoPE assigns one dimension to modality distinctions and two dimensions to image spatial coordinates and text positions.
  • The multi-stage training pipeline progresses through pre-training, mid-training, SFT, GRPO, and DPO for text-to-image, while editing initializes from a T2I development checkpoint.
  • The built-in prompt-refinement solution reuses Qwen2.5-VL to avoid external-service dependencies while supporting high-fidelity image generation.

4 Model Training

LongCat-Image uses a progressive, multi-stage training pipeline that combines resolution curricula, intensive data curation, specialized rendering strategies, and reward-guided reinforcement learning. The training design balances visual quality, adaptability, text rendering, and optimization efficiency.

  • Training Pipeline: Pre-training, Mid-training, and Post-training form the model’s three-phase training pipeline.Mid-training raises baseline generation quality through aesthetic assessment and human-curated filtering.
  • Training Pipeline: Progressive training moves from 256px through an intermediate 512px stage to dynamic 512px–1024px resolutions with bucket sampling.The intermediate stage avoids computational instability during resolution transitions and accommodates variable aspect ratios.
  • Text Rendering: Over 10 million SynthDoG samples target the long-tail distribution of common and rare Chinese characters.Synthetic text is rendered on simple textures with high typographic diversity to improve character learning.
  • Data Curation: Mid-training uses hierarchical aesthetic, quality, and domain-specific filtering with human verification to produce a high-fidelity corpus.The resulting Developer Version retains plasticity and adaptability for downstream fine-tuning and research.
  • Post-training: SFT improves photorealistic attributes and stylistic fidelity using expert-verified real and synthetic data, model weight averaging, and uniform timestep sampling.Specialized candidate models cover illumination, portraiture, and artistic style, while uniform sampling emphasizes later high-frequency denoising steps.
  • Reinforcement Learning: Reward-guided training combines distortion, AIGC, human-preference, and OCR reward models with DPO, GRPO, and MPO.MPO removes GRPO’s group-relative synchronization bottleneck; DPO reduces bad cases and improves generation robustness.

5 Model Performance

LongCat-Image performs strongly across generation, semantic alignment, text rendering, Chinese character coverage, and perceptual quality evaluations. The reported results emphasize SOTA performance across several benchmarks, while also identifying reduced stability for multi-character Chinese sequences.

  • General Generation: LongCat-Image shows superior GenEval performance for attribute binding, quantitative relations, and spatial composition.GenEval contains fine-grained controllability tests for complex compositional constraints and entity attributes.
  • General Generation: LongCat-Image achieves competitive DPG-Bench alignment, ranking closely behind Qwen-Image and Seed4.0 on 1,065 dense prompts.The benchmark tests semantic alignment for verbose and structurally complex captions.
  • General Generation: LongCat-Image reaches SOTA scores among open-source diffusion models on WISE’s 1,000 prompts for semantic comprehension and world knowledge.Evaluation uses an off-the-shelf text encoder for intrinsic prompt enhancement.
  • Text Rendering: LongCat-Image excels on GlyphDraw2’s Complex-Set and attains SOTA performance on CVTG-2K for complex bilingual and multi-region text rendering.GlyphDraw2 tests bilingual design text and Chinese-character coverage, while CVTG-2K evaluates English text placement in real-world scenes.
  • Chinese Character Rendering: On ChineseWord’s 8,105-prompt benchmark, LongCat-Image outperforms existing models significantly, but multi-character generation shows reduced stability.The authors attribute this limitation primarily to insufficient real-world textual training data.
  • Text Rendering: LongCat-Image delivers SOTA-level performance on Poster&SceneBench for poster typography and text integrated into natural scenes.The 500-prompt benchmark covers both flat poster layouts and textured, variably lit real-world environments.
  • Human Evaluation: LongCat-Image is reported to outperform Qwen-Image and slightly exceed Seedream 4.0 in Visual Realism, despite a marginal aesthetics gap versus Qwen-Image.Qualitative comparisons also report robust performance across alignment, plausibility, realism, aesthetics, and text rendering.

6 Image Editing

LongCat-Image-Edit adapts a text-to-image foundation model for image editing and targets broad, high-quality editing data. Its training resources combine curated datasets, synthetic pairs, video-derived transitions, and web-scale interleaved corpora, although long-tail coverage remains limited.

  • LongCat-Image-Edit extends a text-to-image foundation model and achieves SOTA performance among open-source models.
  • Data Construction: The editing dataset combines open-source datasets, synthetic pipelines, video sequences, and interleaved web corpora.Instructions are extensively rewritten to cover simple descriptions through complex reasoning.
  • Synthetic Data: Synthetic pipelines generate pairs for object manipulation, style transfer, background alteration, and reference-based generation.MLLMs craft instructions, expert models generate targets, and traditional image-processing algorithms handle low-level adjustments.
  • Video Data: Video sequences provide editing pairs with significant but coherent structural changes that synthetic methods often reproduce poorly.The pipeline uses multimodal models and optical flow to identify objects, quantify frame changes, and extract keyframe pairs.
  • Interleaved Corpus: Web-scale interleaved corpora enrich long-tail editing diversity, but the curated dataset remains limited because mining valid samples is resource-intensive.The extracted pairs undergo filtering and multimodal-assisted instruction rewriting.

6.2 Model Design

The editing model adds an image-conditioning branch that encodes reference images alongside noised latents while preserving spatial alignment.

  • Reference images are encoded into VAE latents and distinguished from noised latents through the first dimension of 3D RoPE embeddings.
  • The remaining 3D RoPE dimensions preserve spatial alignment between reference-image tokens and noised latents.
  • Reference tokens are concatenated with the model’s latent sequence to provide image conditioning during editing.

6.3 Model Training

LongCat-Image-Edit uses a progressive multi-stage curriculum, joint T2I-editing training, strict alignment filtering, and DPO to improve fidelity, adherence, and robustness.

  • The training framework progresses through Pre-training, SFT, and DPO stages to enhance generated-image resolution and visual fidelity.
  • Pre-training: Training begins at 512×512 with noisy data and advances to 1024×1024 with high-quality data while mixing editing and T2I data.
  • SFT: Strict human-in-the-loop filtering targets structural alignment between source and edited images during SFT.
  • SFT: Jointly training the filtered editing corpus with high-quality T2I SFT data improves instruction adherence and aesthetic quality.
  • DPO: DPO follows SFT with manually annotated preferences to align outputs with human aesthetic standards and reduce structural artifacts.

6.4 Discussion

The authors found that unifying T2I and editing degraded T2I photorealism because editing pre-training relied heavily on synthetic data. They therefore separated the models while identifying larger interleaved corpora as a possible route to future unification.

  • A unified T2I-editing model was initially pursued to exploit task synergies and reduce deployment costs.
  • Heavy reliance on synthetic editing data degraded T2I photorealism relative to training solely on real data.
  • The authors separated the models and attribute the incompatibility to data quality rather than an architectural flaw.
  • They suggest that large-scale interleaved corpora could enable future unification without sacrificing generation quality.

6.5 Model Performance

LongCat-Image-Edit is evaluated across quantitative benchmarks, human comparisons, and qualitative editing scenarios. It achieves strong open-source benchmark performance, preserves consistency through complex edits, and remains behind some commercial systems in human evaluation.

  • Quantitative Evaluation: 1,464 bilingual editing pairs across 15 fine-grained task categories define the comprehensive CEdit-Bench evaluation suite.CEdit-Bench expands existing benchmarks to improve task coverage and granularity.
  • Quantitative Evaluation: LongCat-Image-Edit achieves state-of-the-art performance among open-source models on CEdit-Bench.Evaluation uses Semantic Consistency, Perceptual Quality, and Overall Score with GPT-4o automation.
  • Quantitative Evaluation: LongCat-Image-Edit outperforms competitors on ImgEdit-Bench and delivers top-tier bilingual instruction-following performance on GEdit-Bench.ImgEdit-Bench also evaluates editing quality and detail preservation.
  • Human Evaluation: Human evaluation shows LongCat-Image-Edit surpasses Qwen-Image-Edit and FLUX.1 Kontext in comprehensive quality and consistency, but trails Nano Banana and Seedream 4.0.The win rate is calculated as (#Win + 0.5 × #Tie)/#Total.
  • Qualitative Results: Qualitative tests show consistent semantic and structural preservation during multi-turn and compound editing sequences.The evaluation includes sequential editing and complex prompts containing multiple requested operations.
  • Qualitative Results: The model handles fine-grained portrait, viewpoint, lighting, interaction, and general image edits while preserving identity and background details.Across general editing dimensions, it consistently outperforms Qwen-Image-Edit and FLUX.1 Kontext, and sometimes exceeds commercial counterparts.

7 Conclusion

The paper presents LongCat-Image as a 6B-parameter diffusion framework designed to balance high-fidelity generation with inference efficiency. It reports strong generation, Chinese text rendering, and open-source editing performance, alongside an open development ecosystem.

  • LongCat-Image is a 6B-parameter diffusion framework using efficient architectural design and refined training methodologies.

8 Contributions and Acknowledgments

The acknowledgments define contributors as people responsible for primary technical and infrastructure work across the LongCat-Image development cycle.

  • Contributors undertook primary responsibilities in data curation, model design, model training, and related infrastructure.
Loading 2512.07584v1…