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
核心贡献 · Key contributions
提出 HunyuanImage 3.0，一个在自回归框架中统一理解与生成的原生多模态模型。 Proposes HunyuanImage 3.0, a native multimodal model unifying understanding and generation in an autoregressive framework.
引入原生思维链机制，实现图像生成的自动化推理。 Introduces a native Chain-of-Thought schema for automated reasoning in image generation.
开发了包含分层描述和事实锚定的全面数据整理流程。 Develops a comprehensive data curation pipeline with hierarchical captioning and factual grounding.
以 80B MoE 模型（13B 激活参数）实现开源图像生成的最先进水平。 Achieves state-of-the-art open-source image generation with an 80B MoE model (13B activated).
设计了结合 SFT、DPO、MixGRPO、SRPO 和 ReDA 的新型后训练流程。 Designs a novel post-training pipeline combining SFT, DPO, MixGRPO, SRPO, and ReDA.
提出 SSAE，一种使用 LLM 和 MLLM 的结构化语义对齐评估指标。 Proposes SSAE, a structured semantic alignment evaluation metric using LLMs and MLLMs.
局限 · Limitations
仅发布文生图功能，图生图等其他能力尚未开放。 Only text-to-image generation is released; image-to-image and other capabilities are not yet available.
评估依赖自定义基准（SSAE），可能无法完全泛化到所有场景。 Evaluation relies on a custom benchmark (SSAE) which may not fully generalize to all scenarios.
模型训练和推理需要大量算力资源。 The model requires significant computational resources for training and inference.
训练数据中的潜在偏差可能影响对代表性不足领域的生成质量。 Potential biases in training data may affect generation quality for underrepresented domains.
思维链推理通过小数据集微调激活，限制了其鲁棒性。 The Chain-of-Thought reasoning is activated via fine-tuning on a small dataset, limiting its robustness.