January 26, 2026: 🚀 HunyuanImage-3.0-Instruct-Distil - Distilled checkpoint for efficient deployment (8 steps sampling recommended).
January 26, 2026: 🎉 HunyuanImage-3.0-Instruct - Release of Instruct (with reasoning) for intelligent prompt enhancement and Image-to-Image generation for creative editing.
HunyuanImage-3.0 is a groundbreaking native multimodal model that unifies multimodal understanding and generation within an autoregressive framework. Our text-to-image and image-to-image model achieves performance comparable to or surpassing leading closed-source models.
HunyuanImage-3.0 Framework
✨ Key Features
🧠 Unified Multimodal Architecture: Moving beyond the prevalent DiT-based architectures, HunyuanImage-3.0 employs a unified autoregressive framework. This design enables a more direct and integrated modeling of text and image modalities, leading to surprisingly effective and contextually rich image generation.
🏆 The Largest Image Generation MoE Model: This is the largest open-source image generation Mixture of Experts (MoE) model to date. It features 64 experts and a total of 80 billion parameters, with 13 billion activated per token, significantly enhancing its capacity and performance.
🎨 Superior Image Generation Performance: Through rigorous dataset curation and advanced reinforcement learning post-training, we've achieved an optimal balance between semantic accuracy and visual excellence. The model demonstrates exceptional prompt adherence while delivering photorealistic imagery with stunning aesthetic quality and fine-grained details.
💭 Intelligent Image Understanding and World-Knowledge Reasoning: The unified multimodal architecture endows HunyuanImage-3.0 with powerful reasoning capabilities. It under stands user's input image, and leverages its extensive world knowledge to intelligently interpret user intent, automatically elaborating on sparse prompts with contextually appropriate details to produce superior, more complete visual outputs.
🚀 Usage
📦 Environment Setup
🐍 Python: 3.12+ (recommended and tested)
⚡ CUDA: 12.8
📥 Install Dependencies
bash
1# 1. First install PyTorch (CUDA 12.8 Version)2pip installtorch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 --index-url https://download.pytorch.org/whl/cu128
34# 2. Install tencentcloud-sdk for Prompt Enhancement (PE) only for HunyuanImage-3.0 not HunyuanImage-3.0-Instruct5pip install -i https://mirrors.tencent.com/pypi/simple/ --upgrade tencentcloud-sdk-python
67# 3. Then install other dependencies8pip install -r requirements.txt
For up to 3x faster inference, install these optimizations:
bash
1# FlashInfer for optimized moe inference. v0.5.0 is tested.2pip install flashinfer-python==0.5.0
💡Installation Tips: It is critical that the CUDA version used by PyTorch matches the system's CUDA version.
FlashInfer relies on this compatibility when compiling kernels at runtime.
GCC version >=9 is recommended for compiling FlashAttention and FlashInfer.
⚡ Performance Tips: These optimizations can significantly speed up your inference!
💡Notation: When FlashInfer is enabled, the first inference may be slower (about 10 minutes) due to kernel compilation. Subsequent inferences on the same machine will be much faster.
HunyuanImage-3.0 (Text-to-image)
🔥 Quick Start with Transformers
1️⃣ Download model weights
bash
1# Download from HuggingFace and rename the directory.2# Notice that the directory name should not contain dots, which may cause issues when loading using Transformers.3hf download tencent/HunyuanImage-3.0 --local-dir ./HunyuanImage-3
2️⃣ Run with Transformers
python
1from transformers import AutoModelForCausalLM
23# Load the model4model_id ="./HunyuanImage-3"5# Currently we can not load the model using HF model_id `tencent/HunyuanImage-3.0` directly 6# due to the dot in the name.78kwargs =dict(9 attn_implementation="sdpa",# Use "flash_attention_2" if FlashAttention is installed10 trust_remote_code=True,11 torch_dtype="auto",12 device_map="auto",13 moe_impl="eager",# Use "flashinfer" if FlashInfer is installed14)1516model = AutoModelForCausalLM.from_pretrained(model_id,**kwargs)17model.load_tokenizer(model_id)1819# generate the image20prompt ="A brown and white dog is running on the grass"21image = model.generate_image(prompt=prompt, stream=True)22image.save("image.png")
1# Download from HuggingFace2hf download tencent/HunyuanImage-3.0 --local-dir ./HunyuanImage-3
3️⃣ Run the Demo
The Pretrain Checkpoint does not automatically rewrite or enhance input prompts, for optimal results currently, we recommend community partners to use deepseek to rewrite the prompts. You can go to Tencent Cloud to apply for an API Key.
bash
1# Without PE2exportMODEL_PATH="./HunyuanImage-3"3python3 run_image_gen.py \4 --model-id $MODEL_PATH\5 --verbose 1\6 --prompt "A brown and white dog is running on the grass"\7 --bot-task image \8 --image-size "1024x1024"\9 --save ./image.png \10 --moe-impl flashinfer
1112# With PE13exportDEEPSEEK_KEY_ID="your_deepseek_key_id"14exportDEEPSEEK_KEY_SECRET="your_deepseek_key_secret"15exportMODEL_PATH="./HunyuanImage-3"16python3 run_image_gen.py \17 --model-id $MODEL_PATH\18 --verbose 1\19 --prompt "A brown and white dog is running on the grass"\20 --bot-task image \21 --image-size "1024x1024"\22 --save ./image.png \23 --moe-impl flashinfer \24 --rewrite 125
4️⃣ Command Line Arguments
Arguments
Description
Recommended
--prompt
Input prompt
(Required)
--model-id
Model path
(Required)
--attn-impl
Attention implementation. Either sdpa or flash_attention_2.
sdpa
--moe-impl
MoE implementation. Either eager or flashinfer
flashinfer
--seed
Random seed for image generation
None
--diff-infer-steps
Diffusion infer steps
50
--image-size
Image resolution. Can be auto, like 1280x768 or 16:9
auto
--save
Image save path.
image.png
--verbose
Verbose level. 0: No log; 1: log inference information.
0
--rewrite
Whether to enable rewriting
1
🎨 Interactive Gradio Demo
Launch an interactive web interface for easy text-to-image generation.
1️⃣ Install Gradio
pip install gradio>=4.21.0
2️⃣ Configure Environment
bash
1# Set your model path2exportMODEL_ID="path/to/your/model"34# Optional: Configure GPU usage (default: 0,1,2,3)5exportGPUS="0,1,2,3"67# Optional: Configure host and port (default: 0.0.0.0:443)8exportHOST="0.0.0.0"9exportPORT="443"
3️⃣ Launch the Web Interface
Basic Launch:
sh run_app.sh
With Performance Optimizations:
bash
1# Use both optimizations for maximum performance2sh run_app.sh --moe-impl flashinfer --attn-impl flash_attention_2
4️⃣ Access the Interface
🌐 Web Interface: Open your browser and navigate to http://localhost:443 (or your configured port)
Latest Version (Image-to-image & Text-image-to-image)
HunyuanImage-3.0-Instruct (Instruction reasoning and Image-to-image generation, including editing and multi-image fusion)
🔥 Quick Start with Transformers
1️⃣ Download model weights
bash
1# Download from HuggingFace and rename the directory.2# Notice that the directory name should not contain dots, which may cause issues when loading using Transformers.3hf download tencent/HunyuanImage-3.0-Instruct --local-dir ./HunyuanImage-3-Instruct
2️⃣ Run with Transformers
python
1from transformers import AutoModelForCausalLM
23# Load the model4model_id ="./HunyuanImage-3-Instruct"5# Currently we can not load the model using HF model_id `tencent/HunyuanImage-3.0-Instruct` directly 6# due to the dot in the name.78kwargs =dict(9 attn_implementation="sdpa",10 trust_remote_code=True,11 torch_dtype="auto",12 device_map="auto",13 moe_impl="eager",# Use "flashinfer" if FlashInfer is installed14 moe_drop_tokens=True,15)1617model = AutoModelForCausalLM.from_pretrained(model_id,**kwargs)18model.load_tokenizer(model_id)1920# Image-to-Image generation (TI2I)21prompt ="基于图一的logo，参考图二中冰箱贴的材质，制作一个新的冰箱贴"2223input_img1 ="./assets/demo_instruct_imgs/input_1_0.png"24input_img2 ="./assets/demo_instruct_imgs/input_1_1.png"25imgs_input =[input_img1, input_img2]2627cot_text, samples = model.generate_image(28 prompt=prompt,29 image=imgs_input,30 seed=42,31 image_size="auto",32 use_system_prompt="en_unified",33 bot_task="think_recaption",# Use "think_recaption" for reasoning and enhancement34 infer_align_image_size=True,# Align output image size to input image size35 diff_infer_steps=50,36 verbose=237)3839# Save the generated image40samples[0].save("image_edit.png")
Custom system prompt. Used when --use-system-prompt is custom
None
--bot-task
Task type. image for direct generation; auto for text; recaption for re-write->image; think_recaption for think->re-write->image
think_recaption
--save
Image save path
image.png
--verbose
Verbose level
2
--reproduce
Whether to reproduce the results
True
--infer-align-image-size
Whether to align the target image size to the src image size
True
--max_new_tokens
Maximum number of new tokens to generate
2048
--use-taylor-cache
Use Taylor Cache when sampling
False
5️⃣ For fewer Sampling Steps
We recommend using the model HunyuanImage-3.0-Instruct-Distil with --diff-infer-steps 8, while keeping all other recommended parameter values unchanged.
bash
1# Download HunyuanImage-3.0-Instruct-Distil from HuggingFace2hf download tencent/HunyuanImage-3.0-Instruct-Distil --local-dir ./HunyuanImage-3-Instruct-Distil
34# Run the demo with 8 steps to samples5exportMODEL_PATH="./HunyuanImage-3-Instruct-Distil"6bash run_demo_instruct_Distil.sh
🧱 Models Cards
📊 Evaluation
Evaluation of HunyuanImage-3.0-Instruct
👥 GSB (Human Evaluation)
We adopted the GSB (Good/Same/Bad) evaluation method commonly used to assess the relative performance between two models from an overall image perception perspective. In total, we utilized 1,000+ single- and multi-images editing cases, generating an equal number of image samples for all compared models in a single run. For a fair comparison, we conducted inference only once for each prompt, avoiding any cherry-picking of results. When comparing with the baseline methods, we maintained the default settings for all selected models. The evaluation was performed by more than 100 professional evaluators.
Human Evaluation with Other Models
Evaluation of HunyuanImage-3.0 (Text-to-Image)
🤖 SSAE (Machine Evaluation)
SSAE (Structured Semantic Alignment Evaluation) is an intelligent evaluation metric for image-text alignment based on advanced multimodal large language models (MLLMs). We extracted 3500 key points across 12 categories, then used multimodal large language models to automatically evaluate and score by comparing the generated images with these key points based on the visual content of the images. Mean Image Accuracy represents the image-wise average score across all key points, while Global Accuracy directly calculates the average score across all key points.
Human Evaluation with Other Models
Human Evaluation with Other Models
👥 GSB (Human Evaluation)
We adopted the GSB (Good/Same/Bad) evaluation method commonly used to assess the relative performance between two models from an overall image perception perspective. In total, we utilized 1,000 text prompts, generating an equal number of image samples for all compared models in a single run. For a fair comparison, we conducted inference only once for each prompt, avoiding any cherry-picking of results. When comparing with the baseline methods, we maintained the default settings for all selected models. The evaluation was performed by more than 100 professional evaluators.
Human Evaluation with Other Models
🖼️ Showcase
Our model can follow complex instructions to generate high‑quality, creative images.
HunyuanImage 3.0 Demo
For text-to-image showcases in HunyuanImage-3.0, click the following links:
HunyuanImage-3.0-Instruct demonstrates powerful capabilities in intelligent image generation and editing. The following showcases highlight its core features:
🧠 Intelligent Visual Understanding and Reasoning (CoT Think): The model performs structured thinking to analyze user's input image and prompt, expand user's intent and editing tasks into a stucture, comprehnsive instructions, and leading to a better image generation and editing performance.
breaking down complex prompts and editing tasks into detailed visual components including subject, composition, lighting, color palette, and style.
✏️ Prompt Self-Rewrite: Automatically enhances sparse or vague prompts into professional-grade, detail-rich descriptions that capture the user's intent more accurately.
🎨 Text-to-Image (T2I): Generates high-quality images from text prompts with exceptional prompt adherence and photorealistic quality.
🖼️ Image-to-Image (TI2I): Supports creative image editing, including adding elements, removing objects, modifying styles, and seamless background replacement while preserving key visual elements.
🔀 Multi-Image Fusion: Intelligently combines multiple reference images (up to 3 inputs) to create coherent composite images that integrate visual elements from different sources.
Showcase 1: Detailed Thought and Reasoning Process
HunyuanImage-3.0-Instruct Showcase 1
Showcase 2: Creative T2I Generation with Complex Scene Understanding
Prompt: 3D 毛绒质感拟人化马，暖棕浅棕肌理，穿藏蓝西装、白衬衫，戴深棕手套；疲惫带期待，坐于电脑前，旁置印 "HAPPY AGAIN" 的马克杯。橙红渐变背景，配超大号藏蓝粗体 "马上下班"，叠加米黄 "Happy New Year" 并标 "(2026)"。橙红为主，藏蓝米黄撞色，毛绒温暖柔和。
HunyuanImage-3.0-Instruct Showcase 2
Showcase 3: Precise Image Editing with Element Preservation
HunyuanImage-3.0-Instruct Showcase 3
Showcase 4: Style Transformation with Thematic Enhancement
HunyuanImage-3.0-Instruct Showcase 4
Showcase 5: Advanced Style Transfer and Product Mockup Generation
HunyuanImage-3.0-Instruct Showcase 5
Showcase 6: Multi-Image Fusion and Creative Composition
HunyuanImage-3.0-Instruct Showcase 6
📚 Citation
If you find HunyuanImage-3.0 useful in your research, please cite our work:
bibtex
1@article{cao2025hunyuanimage,
2 title={HunyuanImage 3.0 Technical Report},
3 author={Cao, Siyu and Chen, Hangting and Chen, Peng and Cheng, Yiji and Cui, Yutao and Deng, Xinchi and Dong, Ying and Gong, Kipper and Gu, Tianpeng and Gu, Xiusen and others},
4 journal={arXiv preprint arXiv:2509.23951},
5 year={2025}
6}
🙏 Acknowledgements
We extend our heartfelt gratitude to the following open-source projects and communities for their invaluable contributions:
