Hunyuan Image 2.1 - Tencent's open-source text-based image model

What is Mixed Graph 2.1?

HunyuanImage 2.1 is an open-source text-to-image generation model launched by Tencent. It supports native 2K resolution and possesses powerful capabilities for understanding complex semantics, accurately generating scene details, facial expressions, and actions. The model supports both Chinese and English input and can generate images in various styles, such as comics and figurines, while maintaining stable control over text and details within the images. Based on dual-channel text encoders and high-compression VAEs, the model significantly improves training and inference efficiency. The model is now open-source, facilitating research and development of derivative models. Users can experience the model's generation capabilities online through Tencent's Hunyuan Large Model platform.

Main functions of mixed image 2.1

• Complex semantic understandingIt supports complex semantic ultra-long prompts of up to 1000 tokens, and can accurately generate scene details, character expressions and actions of multiple objects.
• Text and detail controlIt supports fine-grained control over text in images, allowing text to blend naturally with the image and reducing text errors.
• Style diversityIt supports generating images in various styles, such as realistic figures, comics, and vinyl figures, while also possessing high aesthetic appeal.
• High-resolution generationNatively supports 2K resolution image generation, suitable for high-fidelity design needs.

Technical Principles of Mixed-Element Image 2.1

• Dual-channel text encoderUsing a general text encoder and a text encoder, we can better understand scene descriptions, character actions, and detailed requirements. The MLLM module enhances text-image alignment capabilities, and the ByT5 model improves text generation expressiveness.
• Structured CaptionStructured captions provide multi-layered semantic information, significantly improving the model's responsiveness to complex semantics. The introduction of an OCR agent and IP RAG addresses the shortcomings of general VLM captioners in describing dense text and world knowledge.
• High compression ratio VAEUsing a VAE with a 32x compression ratio significantly reduces the computational cost of model training and inference. Utilizing dinov2 alignment and repa loss reduces training difficulty and improves model generation efficiency.
• Two-stage intensive post-trainingTraining: Based on two stages of training: Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL). A self-developed Reward Distribution Alignment reinforcement learning algorithm was used, innovatively introducing high-quality images as chosen samples, significantly improving model performance.
• Multi-resolution trainingIt supports multi-resolution repa loss, accelerates model convergence, and improves the clarity and texture of generated images.

Project address for Hunyuan Image 2.1

• Project official websitehttps://hunyuan.tencent.com/image
• GitHub repository: https://github.com/Tencent-Hunyuan/HunyuanImage-2.1
• HuggingFace model library: https://huggingface.co/tencent/HunyuanImage-2.1

Application scenarios of mixed-element image 2.1

• Creative illustration and designDesigners can generate high-fidelity creative illustrations, such as illustrations with specific styles, scenes and characters based on descriptions, for use in publications such as books and magazines.
• Poster and Packaging DesignIt can create posters and packaging designs that include promotional slogans in both Chinese and English, accurately presenting the integration of text and images, and improving design efficiency and quality.
• Comic creationIt supports the generation of complex four-panel comics and serialized comics, allowing creators to quickly transform their ideas into coherent comic stories and enrich their creative content.
• Game art asset generationIt supports the generation of art assets such as characters, scenes, and props in games, helping game developers quickly build game worlds and reduce development costs.
• Education and learning supportIn the field of education, it is used to generate teaching illustrations, recreate historical scenes, etc., to help students understand knowledge more intuitively and enhance their interest in learning.