HunyuanImage 3.0 Technical Report

TL;DR

HunyuanImage 3.0 is a native multimodal model that unifies image understanding and generation in a single autoregressive backbone (Hunyuan-A13B, an 80B-total / 13B-activated MoE LLM), with text tokens predicted by next-token prediction and image tokens predicted by an inline diffusion head over VAE latents. The image-generation module is open-weights and is, at release, the largest open-source image generative model. Two specific architectural choices stand out: a Generalized Causal Attention mask that mixes 1D causal text attention with full attention inside image segments, and a single 16× VAE that replaces the conventional 8×-VAE + 2× patchifier of DiT-style models. Post-training stacks SFT → DPO → MixGRPO → SRPO + ReDA, and a Chain-of-Thought stage trained on T2T and T2T-and-image data lets the model “think” (rewrite/refine the prompt) before emitting pixels.

Key claims

The backbone is Hunyuan-A13B, a decoder-only MoE LLM with 64 experts and 8 activated per token plus one shared MLP — ≈80B total / 13B activated — extended with a vision encoder and a VAE for joint understanding+generation [§3.1.1, Backbone].
A Generalized Causal Attention mask is used: text tokens attend only to preceding multimodal tokens; image tokens attend to all preceding multimodal tokens and to all other image tokens in the same image segment [§3.1.2].
A single 16× downsampling VAE projecting into a 32-dim latent space replaces the conventional 8×-VAE + 2× patchifier of DiT-style architectures, and is reported to yield superior generation quality [§3.1.1, Image Encoder].
A dual-encoder strategy for conditioning images concatenates VAE latents with vision-encoder features inside the same sequence, supporting interleaved text dialogue, generation, understanding and editing in one continuous context, instead of the segregated VAE-for-gen / ViT-for-understanding split of prior unified models [§3.1.1, Image Encoder].
Image tokens use a Generalized 2D RoPE (2D anisotropic extension of 1D RoPE) and reduce exactly to 1D RoPE in the absence of image tokens, preserving compatibility with the pre-trained LLM [§3.1.3].
The vocabulary is extended with <img_size_*> and <img_ratio_*> tokens (sizes from 256–1024, 33 aspect-ratio anchors 1:4 to 4:1) so the model predicts the appropriate output resolution and aspect ratio from context [§3.1.4].
Pre-training proceeds in 4 progressive stages over T2I, LM, MMU, INTL and CoT tasks, with VAE resolution rising from 256→1024px and ViT input fixed at 512px [§4.1, Table 1].
Post-training is a 5-stage stack: SFT on curated images → DPO (suppress structural distortions from paired SFT-model outputs) → MixGRPO (hybrid ODE-SDE flow GRPO with multi-reward optimization) → SRPO (single-step gradient-guided realism alignment from latent-noise injection) → ReDA (in-house reward-distribution-alignment) [§4.2].
An automatic CoT pipeline trained on (1) T2T reasoning data over real image-gen prompts and (2) T2T-and-image data with explicit reasoning traces lets the model interpret intent, rewrite/refine the prompt, then synthesize the image [§2.3].
Image filtering retained “less than 45%” of an initial 10B-image pool through three filtering stages (technical flaws → learned operators including AIGC and OCR detectors → embedding-cluster dedup), yielding ~5B training images plus 100M+ paired/multi-image samples [§2.1].
Captions are bilingual EN/ZH, hierarchical (4 length tiers + style/cinematography/lighting/atmosphere/composition + factual entities), composed dynamically at train time, and grounded via OCR and Named-Entity agents with bidirectional verification [§2.2].
Automatic + human evaluation of text-image alignment and visual quality is reported to rival prior SOTA (Seedream 4.0, Nano Banana, GPT-Image, HunyuanImage 2.1) [§1, Abstract; §5 partially seen].

Method

HunyuanImage 3.0 is built on top of Hunyuan-A13B (80B-total / 13B-activated decoder-only MoE LLM) extended for vision in both directions. For conditioning on input images, a frozen pretrained vision encoder and a VAE both feed the LLM through dedicated projectors (the VAE path uses a timestep-modulated residual block in the style of DiT AdaLN; the ViT path uses a 2-layer MLP). For generation, image tokens are produced by an inline diffusion head over VAE latents — the same hybrid discrete-text-tokens / continuous-image-tokens scheme as Transfusion and JanusFlow — but unified under a single Generalized Causal Attention mask that lets every image token see every other image token within the same image segment while keeping autoregressive causality across the broader text+image sequence. A Generalized 2D RoPE assigns image tokens 2D positions while collapsing exactly to 1D RoPE on pure-text contexts, so the LLM’s pretrained text capabilities are not disrupted. Output resolution and aspect ratio are predicted by the model itself via special <img_size_*> / <img_ratio_*> vocabulary tokens.

Pre-training spans four progressive stages over T2I, LM, MMU, interleaved text-image, and CoT data with VAE resolution rising from 256px to 1024px while the ViT input stays at 512px. Stage I trains the transformer with ViT frozen; Stage II trains the ViT alone for MMU; Stages III–IV jointly train both at higher resolution with interleaved and reasoning data. Post-training then stacks five stages: supervised fine-tuning on curated high-quality images, DPO on paired good/bad samples generated by the SFT model to suppress structural distortions, MixGRPO (an extension of GRPO to flow models using a hybrid ODE–SDE sampling scheme together with multiple open-source and proprietary reward models, with a refined advantage estimator for faster convergence), SRPO (a single-step gradient-guided RL pass that injects noise into latents and denoises in one step to optimize early-denoising trajectory under differentiable rewards), and finally an in-house Reward Distribution Alignment (ReDA) step that minimizes divergence to a high-reward image distribution. The CoT capability is unlocked by a small reasoning dataset of two kinds: T2T data over real image-generation prompts (covering photoreal, artistic, UI/poster, knowledge-driven and scientific renderings) and T2T-and-image data that pairs curated images with explicit reasoning traces translating user intent into detailed visual specs.

Results

The paper reports that automatic and human evaluation on text-image alignment and visual quality rivals previous SOTA models including Seedream 4.0, Nano Banana, GPT-Image and HunyuanImage 2.1 [§1, Abstract]. Quantitative breakdowns live in §5 (SSAE, GenEval, T2I-CompBench), which the fetched body truncates before fully exposing; the model card and HF page reaffirm the rivals-leading-closed-source positioning. The HunyuanImage 3.0-Instruct variant announced in the source tweet is positioned as an editing-tuned descendant focused on add/remove/modify edits with non-target preservation and on multi-image fusion. At 80B-total / 13B-activated, it is the largest open-source image generative model at release.

Why it’s interesting

Three things are worth flagging for Luma. First, the architectural ablation against the conventional DiT recipe — single 16× VAE in place of 8×-VAE + 2× patchifier, and dual VAE+ViT conditioning instead of mode-split features — is a concrete, scaled datapoint on latent design that complements Scaling Text-to-Image Diffusion Transformers with Representation Autoencoders‘s RAE-vs-VAE comparison; both papers argue the conventional FLUX-style VAE is leaving quality on the table at T2I scale, though they reach for opposite tools (16× VAE here, frozen SigLIP-2 representation encoder there). Second, MixGRPO is the most aggressive deployment to date of GRPO-style RL on a flow-matching image generator at a >10B-active scale, paired with SRPO + ReDA — a richer RL stack than the simpler RLHF/DPO recipes most open image models ship with. Third, the native CoT-before-pixels schema (the model rewrites and refines the prompt internally before emitting image tokens) is a different point in the design space from thinking-with-modalities papers like Video-Thinker: Sparking "Thinking with Videos" via Reinforcement Learning and DiffThinker: Towards Generative Multimodal Reasoning with Diffusion Models: text reasoning gates image synthesis instead of image manipulation being the reasoning trace, and the trick is unlocked with a small post-training dataset rather than an architectural change.