HunyuanImage 3.0 Technical Report

Abstract

HunyuanImage 3.0 is a massive 80-billion parameter Mixture-of-Experts (MoE) model that unifies image understanding and generation into a single native multimodal framework. By integrating Chain-of-Thought (CoT) reasoning and training on a curated 5-billion image dataset, it achieves state-of-the-art performance, rivaling top-tier closed-source models while remaining open-source.

Introduction

Despite the rapid progress in AI-generated content (AIGC), a significant gap remains between open-source and closed-source models. Most existing systems treat image understanding and image generation as separate tasks handled by different architectures. Furthermore, many models struggle with complex prompt-following, spatial reasoning, and "AIGC contamination"—where models are inadvertently trained on low-quality AI-generated data, leading to a degradation in visual realism. HunyuanImage 3.0 addresses these gaps by moving away from fragmented pipelines toward a native multimodal architecture.

Unified Multimodal Architecture

The model is built on Hunyuan-A13B, an MoE large language model featuring 80B total parameters, with 13B activated per token to balance high capacity with inference efficiency. It employs a hybrid discrete-continuous strategy: text is handled via standard autoregressive next-token prediction, while image tokens are modeled through a diffusion-based framework integrated directly into the LLM.

Expand

Source, p. 6

Key architectural innovations include:

• Generalized Causal Attention: Text tokens follow strict causal attention, while image tokens use full attention within their own segment to capture global spatial dependencies.
• Generalized 2D RoPE: The authors extended Rotary Position Embeddings (RoPE) to 2D for images while maintaining backward compatibility with the 1D RoPE used for text.
• Native Chain-of-Thought (CoT): Before generating an image, the model performs conceptual refinement and rewriting, mapping abstract user intents to precise visual specifications.

Data Curation and Evaluation

The team implemented a massive data curation pipeline, filtering 10 billion raw images down to a 5-billion high-quality corpus using a three-stage process focusing on technical flaws, aesthetic scoring, and deduplication.

To measure performance, the authors introduced the Structured Semantic Alignment Evaluation (SSAE). This benchmark uses an LLM to parse prompts into 12 fine-grained fields (nouns, actions, styles, etc.) to provide a more objective measure of prompt-following than standard CLIP scores. Specialized datasets were also constructed for CoT, including T2TI (Text-to-Text-to-Image) reasoning traces and TI2TI (Image-to-Image) editing sequences.

Experimental Results

In GSB (Good/Same/Bad) testing, HunyuanImage 3.0 achieved a 14.1% relative win rate over its predecessor (v2.1) and outperformed leading models like Seedream 4.0 and GPT-Image. The model showed superior performance in fine-grained semantic alignment, particularly in complex scenes involving multiple attributes and subjects.

Expand

Source, p. 2

Through distillation (MeanFlow), the model can generate high-quality images in just 4 to 8 steps, significantly reducing the computational cost typically associated with large-scale diffusion models. Analysis of the MoE layers revealed that as the model gets deeper, experts naturally specialize in either image or text modalities, suggesting that MoE is particularly well-suited for multimodal scaling.

Conclusion

HunyuanImage 3.0 represents a significant step toward unified multimodal intelligence. While the current public release focuses primarily on the text-to-image generation module, the underlying architecture is natively multimodal. Future work will expand these capabilities to include full image-to-image and multimodal understanding tasks.

@misc{cao2025hunyuanimage,
  author = {Siyu Cao and Hangting Chen and Peng Chen and Yiji Cheng and Yutao Cui and Xinchi Deng and Ying Dong and Kipper Gong and Tianpeng Gu and Xiusen Gu and Tiankai Hang and Duojun Huang and Jie Jiang and Zhengkai Jiang and Weijie Kong and Changlin Li and Donghao Li and Junzhe Li and Xin Li and Yang Li and Zhenxi Li and Zhimin Li and Jiaxin Lin and Linus and Lucaz Liu and Shu Liu and Songtao Liu and Yu Liu and Yuhong Liu and Yanxin Long and Fanbin Lu and Qinglin Lu and Yuyang Peng and Yuanbo Peng and Xiangwei Shen and Yixuan Shi and Jiale Tao and Yangyu Tao and Qi Tian and Pengfei Wan and Chunyu Wang and Kai Wang and Lei Wang and Linqing Wang and Lucas Wang and Qixun Wang and Weiyan Wang and Hao Wen and Bing Wu and Jianbing Wu and Yue Wu and Senhao Xie and Fang Yang and Miles Yang and Xiaofeng Yang and Xuan Yang and Zhantao Yang and Jingmiao Yu and Zheng Yuan and Chao Zhang and Jian-Wei Zhang and Peizhen Zhang and Shi-Xue Zhang and Tao Zhang and Weigang Zhang and Yepeng Zhang and Yingfang Zhang and Zihao Zhang and Zijian Zhang and Penghao Zhao and Zhiyuan Zhao and Xuefei Zhe and Jianchen Zhu and Zhao Zhong},
  title = {{HunyuanImage 3.0 Technical Report}},
  year = {2025},
  eprint = {2509.23951},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.23951},
}