core
HunyuanImage 3.0 Technical Report
Tencent Hunyuan Team · Tencent · 2025-09 · arXiv:2509.23951
Why it matters
An 80B mixture-of-experts model that unifies understanding and generation in one backbone, with autoregressive text generation and a diffusion head operating on continuous VAE latents for images, closer to Transfusion's hybrid than to pure autoregressive image generation. Includes native chain-of-thought before image output, and was the largest open-sourced unified image foundation model at release.
What this paper does
read: full textBefore this
Prior unified multimodal models split visual features by task, using vision-encoder features for understanding and separate VAE features for generation. Most systems matching state-of-the-art image quality, such as Seedream 4.0, Nano Banana, GPT-Image and Qwen-Image, were closed source, which limited transparency and reproducibility for the research community.
The problem
The paper sets out to build an open-source model whose generation quality is comparable to or better than these leading closed-source systems, while unifying understanding and generation and giving the model an explicit reasoning step before it generates.
The idea
Build one large autoregressive multimodal backbone that models text with standard next-token prediction and interleaves image generation into the same sequence with a diffusion-based prediction head on continuous VAE features, following the hybrid recipe used by Transfusion and JanusFlow. Train the model to produce an intermediate chain-of-thought text span before producing the image so planning happens natively rather than through an external prompt rewriter.
How it works
HunyuanImage 3.0 is built on the Hunyuan-A13B Mixture-of-Experts backbone, over 80 billion total parameters with 13 billion activated per token, using 64 experts with 8 active per token. Images are represented in the continuous latent space of an internal VAE, 32 channels with a 16x spatial downsampling factor, and the fetched text reports no discrete VQ tokens anywhere in the pipeline. Text tokens are predicted with standard autoregressive next-token cross-entropy. Image tokens are produced by a diffusion-based prediction framework attached to the same autoregressive backbone, operating on the VAE's continuous latent features, in the manner the paper states follows Transfusion and JanusFlow; the fetched text does not specify whether the diffusion loss regresses noise, velocity, or the clean latent, nor whether the VAE itself is frozen or jointly trained. Training proceeds through four progressive stages, moving from a frozen vision encoder with a trainable transformer backbone, to a frozen backbone with a trainable vision encoder, to joint training of both, to full joint training at higher resolution up to 1024 pixel VAE resolution with chain-of-thought data included. Post-training layers reward-based and preference-based refinement on top of the base objective, SFT on curated high-quality images, DPO on preference pairs targeting structural defects, MixGRPO, an online RL framework using hybrid ODE-SDE sampling to optimize aesthetics, SRPO, a gradient-guided online RL step that injects noise priors for single-step denoising refinement, and ReDA, a reward-distribution-alignment step that minimizes divergence from a high-reward distribution.
Evidence
Training used close to 5 billion images retained from an initial pool of over 10 billion after filtering, a retention rate under 45%. On the paper's own SSAE semantic-alignment benchmark the model performs on par with leading models across all measured fine-grained categories. In blind human GSB evaluation on 1,000 prompts judged by over 100 professional evaluators in a single run, HunyuanImage 3.0 wins by a 14.10% relative margin over its own predecessor HunyuanImage 2.1, and by smaller margins over external systems, 1.17% over Seedream 4.0, 2.64% over Nano Banana, and 5.00% over GPT-Image. No FID or other likelihood-style generation metric is reported in the fetched text, and no training compute or token budget is disclosed.
Limitations
The authors concede that only the text-to-image capability is released at this point, and describe image-to-image editing training as ongoing with release promised in the near future. The fetched text gives no explicit failure-mode or ablation discussion, and omits the exact diffusion training objective and the VAE's training status, which limits how precisely its generative mechanism can be compared to other hybrid autoregressive-diffusion models.
Why it matters
At over 80 billion total parameters with weights released, this is presented as the largest open-source image generative model at the time of release. It demonstrates that a Transfusion/JanusFlow-style hybrid, autoregressive text combined with diffusion image tokens in one backbone, scales to match closed frontier systems on human preference while also supporting native chain-of-thought before generation.
Abstract, in the authors' own words
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
Research line
Design-axis choices
Conditioning & control
Method note — the shared flow-matching interpolation
Every flow-matching / rectified-flow paper in this atlas trains toward a straight-line path between a noise sample x₀ and a data sample x₁:
Builds on
Nothing recorded yet.
Built on by
Nothing recorded yet.