HunyuanImage 3.0 Technical Report

• Against leading systems that are predominantly closed-source, the report releases HunyuanImage 3.0’s code and weights to support community exploration and an image-generation ecosystem.Source: §1 · pp. 1, 3
• The model builds on Hunyuan-A13B, an MoE LLM with over 80 billion total parameters and 13 billion activated per inference token.A pre-trained vision encoder and VAE project image features into a joint embedding space compatible with the LLM’s word embeddings.Source: §1 · p. 1
• Autoregressive next-token prediction supports image understanding, while diffusion-based modeling of VAE image features enables generation; the framework also incorporates Chain-of-Thought training and inference.Source: §1 · p. 1
• Evaluations on text-image alignment and visual quality show HunyuanImage 3.0 rivals previous state-of-the-art models, including Seedream 4.0, Nano Banana, GPT-Image, and HunyuanImage 2.1.Source: §1 · p. 3

2.1 Data Filtering

• From over 10 billion raw images, three-stage filtering retained less than 45%, producing nearly 5 billion training images and over 100 million image pairs and multi-image samples.The paired and multi-image corpus supports learning interleave relationships.Source: §2.1 · p. 3
• Filtering removed technically flawed and duplicate images, then used learned objective detectors and clarity and aesthetics scoring to curate the data.Detectors target watermarks, logos, extensive text, collages, borders, and AI-generated content; sources with high proportions of AI-generated images were also removed. A unified threshold filtered unqualified images, with genre-specific thresholds selecting data for later training.Source: §2.1 · p. 3
• Image-pair data came from similarity-bearing image clusters and video segments, with relation, complexity, camera-motion, object, segmentation, and blur filters refining selections.The relation operator retained pairs with significant editing relationships, and the complexity model filtered overly complex images. Video mining isolated unified scenes and canonical transformation relationships; the resulting multi-image data included internet image-text sequences and extracted video frames.Source: §2.1 · pp. 3–4

2.2 Image Captioning

• A bilingual English/Chinese hierarchical schema organizes captions by descriptive length, stylistic attributes, and named entities, while compositional synthesis varies fields and patterns.Generated captions range from about 30 to 1,000 words; the schema also provides fine-grained generative control and a structural basis for data synthesis.Source: §2.2 · p. 4
• OCR and named-entity agents supply auxiliary facts that captions must pass bidirectional verification, while a paired-image captioner describes foreground and background changes for editing instructions.The image-difference captioner uses image pairs, their captions, and corresponding two-frame video as input.Source: §2.2 · pp. 4–5

2.3 Reasoning Dataset Construction

• Specialized T2T, T2TI, and TI2TI datasets train stepwise reasoning from prompts to image specifications and outputs, including complex compositional editing.T2T covers diverse generation prompts; T2TI pairs images and captions with annotated reasoning traces; TI2TI includes source images, complex editing instructions, atomic-operation traces, and ground-truth edits.Source: §2.3 · p. 5

3.1 Native Multimodal Model

• Text tokens use autoregressive next-token prediction, while image tokens use diffusion-based prediction in a unified multimodal framework.The model brings together language modeling, image understanding, and image generation.Source: §3.1 · p. 5

3.1.1 Architecture

• The Hunyuan-A13B decoder-only backbone has over 80 billion total parameters, with 64 experts and 8 activated per token for approximately 13 billion active parameters.It also includes one shared MLP.Source: §3.1.1 · p. 6
• A 16x-downsampling VAE maps pixels to 32-dimensional latents, and concatenated VAE and vision-encoder features support generation and understanding in one sequence.The report describes the single VAE as a simpler, more effective alternative to prior 8x-downsampling VAE-plus-patchification approaches, yielding superior image generation quality. The dual-encoder strategy supports interleaved dialogue, generation, understanding, and editing without switching pipelines.Source: §3.1.1 · p. 6

3.1.2 Generalized Causal Attention

• Generalized Causal Attention limits text tokens to preceding multimodal tokens while image tokens attend to preceding tokens and successive image tokens within the same image segment.For training sequences with multiple generated images, images in context are masked from subsequent tokens; at inference, generated images become conditional images, so the canonical mask suffices.Source: §3.1.2 · pp. 6–7

3.1.3 Position Embedding

• Generalized 2D RoPE assigns image tokens 2D positions while retaining standard 1D RoPE for text, reducing exactly to 1D RoPE when no image tokens are present.This preserves compatibility with the pretrained LLM; positions after generated images are shifted during training to align with inference positions in sequences containing multiple generated images.Source: §3.1.3 · p. 8

3.1.4 Automatic Resolution

• Automatic resolution uses predicted size-anchor and aspect-ratio tokens, with ratios spanning 1:4 to 4:1, to set image shape and enable 2D RoPE.The model predicts these tokens from context, and users can provide cues such as “3:4” or “vertical” to guide aspect ratio.Source: §3.1.4 · p. 8

4.1 Pre-training

• Four-stage pre-training progressively filters data and raises VAE image resolution while keeping ViT resolution fixed, then instruction tuning jointly optimizes T2I, LM, and CoT data.The stages train T2I, LM, and MMU, fine-tune the ViT and aligner on MMU, jointly train ViT and Transformer with higher-resolution and interleaved data, then add high-resolution MMU and reasoning data; reasoning tokens are modeled by autoregressive next-token prediction. Image aspect ratios are preserved, and the ViT input remains fixed at 512 pixels in the final stage.Source: §4.1 · pp. 8–9, Table 1 · p. 9

4.2 Post-training

• Post-training combines SFT, DPO, MixGRPO, SRPO, and ReDA to refine instruction following, reduce artifacts and distortions, and improve alignment, realism, and aesthetics.SFT uses curated text-to-image and editing data, reasoning examples, and progressively higher-fidelity samples; DPO uses paired data followed by a high-quality subset. MixGRPO applies hybrid ODE–SDE sampling with reward models, while balanced multi-task reward training targets editing and identity preservation. SRPO injects a noise prior into latent features and uses positive and negative text-guided rewards; ReDA aligns generations to a high-reward prior, using reference-to-generation feature transitions when references are available.Source: §4.2 · pp. 9–10

4.3 Distillation

• Reducing NFE to 4–8, the MeanFlow distillation framework scales to the 80B unified multimodal model while preserving competitive performance.The framework mitigates training instability and incorporates trajectory distribution alignment into the MeanFlow objective to improve few-step generation.Source: §4.3 · p. 10

4.4 Pruning

• Compressing HunyuanImage-3.0 from 80B to 20B parameters (75% reduction) with TMP preserves competitive generation quality and enables deployment on a single 24GB RTX 4090 GPU.Tree-structured Mixed-policy Pruning combines heuristic pruning with hierarchical local mixed-policy distillation; memory-efficient inference optimizations support the single-GPU deployment.Source: §4.4 · p. 10

5.1 SSAE

• SSAE addresses benchmark prompt limitations with 500 diverse prompts, parsed into 3,500 key points across 12 fields and checked by an LLM before human rectification.The fixed key points support stable and fair comparisons across text-to-image models.Source: §5.1 · pp. 10–11
• An MLLM with Chain-of-Thought reasoning scores image-to-prompt matches from 0 to 1, yielding field-specific accuracy, Mean Image Accuracy, and Global Accuracy.SSAE targets the mismatch between automated metrics and human judgment, including errors in spatial relationships and precise attribute binding.Source: §5.1 · pp. 10–11
• SSAE finds HunyuanImage 3.0 performs on par with leading models across all fine-grained fields.Compared with DreamBench++ [45], SSAE uses hierarchical semantic-point parsing and taxonomy, with extensions such as ranking beyond scoring.Source: §5.1 · p. 11, Figure 6 · p. 11

5.2 GSB

• HunyuanImage 3.0 achieves relative win rates of 14.10% over HunyuanImage 2.1, 1.17% over Seedream 4.0, 2.64% over Nano Banana, and 5.00% over GPT-Image.Source: §5.2 · p. 11

5.3.1 Expert Activation Analysis

• Analysis of 1,000 prompts finds experts increasingly specialize by modality across layers, with deeper layers showing higher KL divergence and more dispersed activation distributions.The report suggests this specialization may help MoE multimodal modeling by distributing responsibilities among experts.Source: §5.3.1 · p. 12, Figure 8 · p. 12, §5.2 · p. 12

• The model extends a pre-trained MoE LLM to unify image understanding and generation, using large-scale pre-training on diverse, carefully curated multimodal data.Source: §6 · p. 12
• Native Chain-of-Thought training and inference improve the model’s multimodal performance significantly within its LLM-based framework.Source: §6 · p. 12
• Automatic and human evaluations indicate that HunyuanImage 3.0 rivals existing state-of-the-art models in text-image alignment and visual quality, with strengths in prompt-following, reasoning, concept generalization, and text rendering.Source: §6 · p. 13
• The publicly available release includes text-to-image generation, while training for image-to-image tasks is ongoing and that capability is planned for a future release.Source: §6 · p. 13