PaperlayerView exact revision on arXiv

Source-linked AI summary

HunyuanImage 3.0 Technical Report

Tencent Hunyuan Foundation Model Team

arXiv:2509.23951v3cs.CV

TL;DR

Existing image-generation systems are often closed-source, while standard benchmarks inadequately capture complex instructions and human-perceived utility. HunyuanImage 3.0 unifies multimodal understanding and generation in a native multimodal framework, and its released image-generation module rivals state-of-the-art models in text-image alignment and visual quality.

  • Problem

    State-of-the-art image-generation systems are predominantly closed-source, limiting transparency, while common benchmarks incompletely represent complex instructions and human-perceived utility.

  • Method

    HunyuanImage 3.0 extends a pretrained MoE language model into a unified multimodal architecture with autoregressive text modeling, diffusion-based image modeling, native Chain-of-Thought reasoning, and multi-stage post-training.

  • Results

    HunyuanImage 3.0 uses over 80 billion total parameters with 13 billion activated per token and rivals existing state-of-the-art models in text-image alignment and visual quality under automatic and human evaluation.

  • Takeaways & Limitations

    The publicly released code and weights provide an open-source foundation model for exploring multimodal image-generation ideas.

  • Takeaways & Limitations

    Benchmark metrics can poorly proxy human assessment when images contain spatial-relation or precise attribute-binding failures.

Abstract

from arXiv · show

We present HunyuanImage 3.0, a native multimodal model that unifies multimodal understanding and generation within an autoregressive framework, with its image generation module publicly available. The achievement of HunyuanImage 3.0 relies on several key components, including meticulous data curation, advanced architecture design, a native Chain-of-Thoughts schema, progressive model pre-training, aggressive model post-training, and an efficient infrastructure that enables large-scale training and inference. With these advancements, we successfully trained a Mixture-of-Experts (MoE) model comprising over 80 billion parameters in total, with 13 billion parameters activated per token during inference, making it the largest and most powerful open-source image generative model to date. We conducted extensive experiments and the results of automatic and human evaluation of text-image alignment and visual quality demonstrate that HunyuanImage 3.0 rivals previous state-of-the-art models. By releasing the code and weights of HunyuanImage 3.0, we aim to enable the community to explore new ideas with a state-of-the-art foundation model, fostering a dynamic and vibrant multimodal ecosystem. All open source assets are publicly available at https://github.com/Tencent-Hunyuan/HunyuanImage-3.0

1 Introduction

HunyuanImage 3.0 is introduced as an open-source native multimodal model targeting the transparency gap in advanced image generation. Its architecture, training, and evaluations position it as comparable to leading closed-source and state-of-the-art systems.

  • Closed-source state-of-the-art image-generation systems limit transparency despite strong fidelity, semantic alignment, scene synthesis, and editing capabilities.
  • HunyuanImage 3.0 extends an 80-billion-plus-parameter MoE language model with 13 billion activated parameters, a vision encoder, and a VAE for multimodal understanding and generation.The vision encoder and VAE project image features into a joint embedding space compatible with language embeddings.
  • Automatic and human evaluations of text-image alignment and visual quality show results rivaling previous state-of-the-art models, including Seedream 4.0, Nano Banana, GPT-Image, and HunyuanImage 2.1.
  • The report covers data preparation, architecture and algorithms, training strategies, and comparisons with state-of-the-art text-to-image models.

2 Data Preparation

The data pipeline combines large-scale filtering, semantic enrichment, structured captioning, factual verification, and reasoning-data construction. These stages produce diverse training data intended to support controllable, grounded, and coherent image generation.

  • Over 10 billion raw images were filtered to retain less than 45%, producing nearly 5 billion clean, high-quality, diverse images for training.The pipeline also supplemented filtered data with knowledge-augmented, text-related, stylized, and graphic-design collections.
  • Technical filtering removes low-resolution, corrupted, overexposed, underexposed, oversaturated, duplicated, watermarked, collaged, and AI-generated images.
  • Subject-scoring models evaluate clarity and aesthetics using sharpness, noise, dynamic range, color, lighting, shadow, and composition criteria.
  • 2.2 Image Captioning: The image-captioning pipeline uses a bilingual hierarchical schema, compositional synthesis, and specialized agents for factual grounding.The schema organizes narrative detail, stylistic attributes, and named entities, while the pipeline includes an OCR Agent and Named Entity Agent with bidirectional verification.
  • 2.2 Image Captioning: Compositional Caption Synthesis combines schema fields to generate bilingual captions with varied patterns and lengths from about 30 to 1,000 words.
  • 2.3 Reasoning Dataset Construction: Reasoning data includes T2T and T2TI examples that train instruction following, logical inference, conceptual refinement, and translation from user intent to visual specifications.The T2TI corpus pairs images and captions with reasoning traces describing the path from abstract concepts to visual outputs.

3 Model Design

HunyuanImage 3.0 unifies text and image understanding and generation through a hybrid multimodal architecture. Its design combines specialized tokenization, image encoders, attention mechanisms, positional embeddings, and automatic image-shape selection.

  • Native Multimodal Model: Text tokens use autoregressive next-token prediction, while image tokens use diffusion-based prediction within one unified multimodal framework.
  • Backbone: The Hunyuan-A13B backbone uses 64 experts with 8 activated per token, yielding over 80 billion total parameters and approximately 13 billion active parameters.
  • Image Encoder: Conditioned images combine VAE and vision-encoder latent features, supporting generation, understanding, dialogue, and image editing in one continuous context.
  • Generalized Causal Attention: Generalized Causal Attention restricts text to preceding multimodal tokens while allowing image tokens global access within their image segment.
  • Position Embedding: Generalized 2D RoPE assigns two-dimensional positions to image tokens while retaining standard 1D RoPE for text, preserving compatibility with the pretrained language model.
  • Automatic Image Shape: Special image-size and image-ratio tokens let the model infer appropriate output shapes from context or follow explicit aspect-ratio cues.

4 Model Training

HunyuanImage 3.0 uses progressive multi-task pre-training and staged post-training to refine multimodal image generation. Distillation and pruning further reduce inference cost while preserving competitive generation quality.

  • Pre-training: Four progressive pre-training stages support text-to-image, language modeling, multimodal understanding, interleaved text-image modeling, and reasoning tasks.Training filters data from coarse to fine while gradually increasing VAE image resolution and keeping the ViT resolution fixed.
  • Instruction Tuning: Instruction tuning jointly optimizes text-to-image, language-modeling, and Chain-of-Thought data formatted with task-specific templates.
  • Post-training: SFT, DPO, MixGRPO, SRPO, and ReDA progressively improve instruction following, stability, alignment, realism, aesthetics, and image clarity.DPO targets distortions and artifacts, MixGRPO optimizes multiple quality dimensions, SRPO aligns realism and aesthetics, and ReDA optimizes against a high-reward prior.
  • Distillation: MeanFlow distillation reduces the Number of Function Evaluations to 4–8 while preserving competitive model performance.The framework addresses training instability and incorporates trajectory distribution alignment into the distillation objective.
  • Pruning: Tree-structured Mixed-policy Pruning compresses the model from 80B to 20B parameters, a 75% reduction, while maintaining competitive generation quality.With memory-efficient inference optimizations, the compressed model can run on a single 24GB RTX 4090 GPU.

5 Model Performance

The evaluation section introduces SSAE to address benchmark limitations, then reports SSAE and GSB results for HunyuanImage 3.0 against recent models. HunyuanImage 3.0 performs on par with leading models across fine-grained SSAE fields and achieves positive GSB win rates against open- and closed-source systems.

  • SSAE: SSAE addresses limited prompt diversity and automated metrics that may diverge from human judgments of text-image alignment.The benchmark limitations include weak coverage of complex instructions, spatial relationships, and precise attribute binding.
  • SSAE: SSAE parses 500 diverse prompts into 3,500 structured semantic points across 12 fine-grained fields.An LLM-based parser extracts and categorizes the points, while another LLM checks coherence with the original prompts.
  • SSAE: An MLLM scores generated images against prompts and key points, producing field-specific accuracy, Mean Image Accuracy, and Global Accuracy.The scoring uses 0-1 matching with Chain-of-Thought reasoning.
  • SSAE: HunyuanImage 3.0 performs on par with leading models in all fine-grained SSAE fields.The comparison is reported in Figure 6.
  • GSB: Over 100 professional evaluators assessed one generation per prompt across 1,000 prompts without cherry-picking.The protocol compares models from an overall image-perception perspective under their default settings.
  • GSB: 14.10% is HunyuanImage 3.0’s relative GSB win rate over HunyuanImage 2.1, the previously best open-source model.Its relative win rates over Seedream 4.0, Nano Banana, and GPT-Image are 1.17%, 2.64%, and 5.00%, respectively.

6 Conclusion

HunyuanImage 3.0 unifies multimodal understanding and generation in an autoregressive framework, with image generation fine-tuned and post-trained for public release. Evaluations indicate strong image-generation capabilities and performance that rivals existing state-of-the-art models, while image-to-image capability remains under development.

  • Conclusion: HunyuanImage 3.0 unifies multimodal understanding and generation within an autoregressive framework.Its architecture combines autoregressive text-token prediction with diffusion-based image-token prediction.
  • Conclusion: The publicly available model demonstrates prompt-following, reasoning, concept generalization, and text-rendering capabilities for text-to-image generation.Fine-tuning and post-training focus specifically on image generation.
  • Conclusion: Automatic and human evaluations of text-image alignment and visual quality indicate that HunyuanImage 3.0 rivals existing state-of-the-art models.The conclusion reports this result across the evaluated capabilities.
  • Conclusion: The release currently includes text-to-image ability, while training for image-to-image tasks is ongoing.The image-to-image capability is planned for a future release.

7 Project Contributors

The report identifies project sponsors, leadership, and contributors across data, model components, algorithms, pretraining, and post-training. Contributors are listed alphabetically within the stated project roles.

  • Project Roles: Project sponsors are Jie Jiang, Liefeng Bo, Peng Chen, and Yuhong Liu, with Zhao Zhong as project leader.
  • Core Contributors: Data and captioning contributors include Xin Li, Duojun Huang, Xinchi Deng, and Xuefei Zhe.
  • Core Contributors: VAE and model-acceleration contributors include Songtao Liu, Changlin Li, Jianbing Wu, and Peizhen Zhang.
  • Core Contributors: Algorithm and pretraining contributors include Miles Yang, Fanbin Lu, Jian-Wei Zhang, Qixun Wang, Yiji Cheng, and Zijian Zhang.
  • Core Contributors: Post-training contributors include Lucas Wang, Chunyu Wang, Hangting Chen, Hao Wen, Junzhe Li, and additional listed collaborators.
  • Contributors: Additional contributors are listed alphabetically in the report.
Loading 2509.23951v3…