Tencent HunyuanDiT: A Text-to-Image Generation Model by Tencent 252

HunyuanDiT is Tencent's latest open-source text-to-image AI model. It adopts the same DiT (Diffusion With Transformer) architecture as Sora and Stable Diffusion 3, and supports bilingual input and understanding in both Chinese and English. It can understand Chinese and English, and also capture subtle nuances in language, such as context and sentence structure complexity. It also supports interacting with the model through multi-turn dialogue to gradually refine and adjust image content. HunyuanDiT supports generating images at different resolutions.

HunyuanDiT fills the gap in the Chinese-native DiT text-to-image architecture, helping more developers and creators participate in exploring and co-creating visual generation based on the DiT architecture. In addition, HunyuanDiT is also described as the first Chinese-native DiT architecture, supporting bilingual input and understanding in both Chinese and English, with a parameter scale of 1.5 billion.

Features and Optimizations of HunyuanDiT:

• Bilingual Chinese-English text generation support: HunyuanDiT is specially designed to handle bilingual Chinese-English text and can generate high-resolution, high-quality images based on detailed text prompts.

• Long text understanding capability: At the algorithmic level, HunyuanDiT optimizes the model's long text understanding capability, supporting content input of up to 256 characters, while achieving multi-turn image generation and dialogue capabilities to achieve more satisfying results.

• Multimodal visual generation: HunyuanDiT can not only be used for text-to-image, but also as the foundation for multimodal visual generation such as video, which demonstrates its powerful visual generation capabilities.

• Network architecture innovation: It adopts an innovative network architecture that combines bilingual CLIP and multilingual T5 encoders, trained and optimized through a carefully designed data pipeline, supports multi-turn dialogue, and can generate and refine images based on context.

• Open source and commercial use: HunyuanDiT is fully open source and allows free commercial use, providing users with more flexibility and usage scenarios.

• Technology fusion: HunyuanDiT combines the advantages of diffusion models and Transformer architecture to provide powerful visual generation capabilities. This architecture can not only be used for text-to-image, but also for video and other multimodal visual content.

Application Areas of HunyuanDiT:

• Text-to-image generation: The HunyuanDiT model is specifically designed to handle bilingual Chinese-English text and can generate high-resolution, high-quality images based on detailed text prompts.

• Video and 3D content generation: HunyuanDiT can not only support text-to-image, but also serve as the foundation for multimodal visual generation such as video.

• Internal business integration: More than 180 internal Tencent businesses have integrated Tencent Hunyuan, including Tencent Meeting, Tencent Docs, WeCom, Tencent Ads, and WeChat Search.

• Chinese-native support: HunyuanDiT is one of the earliest text-to-image models in the industry to explore and apply large language models combined with the DiT structure, supporting bilingual input and understanding in both Chinese and English, with a parameter scale of 1.5 billion.

• Multi-turn dialogue and image refinement: HunyuanDiT is capable of multi-turn dialogue with users, generating and refining images based on context.

Tencent HunyuanDiT Project Address:https://github.com/Tencent/HunyuanDiT

Tencent HunyuanDiT Project Official Website:https://dit.hunyuan.tencent.com/