Unicom Yuanjing - China Unicom's open-source Chinese native text-to-image model

What is Unicom Yuanjing?

UniT2IXL is a native Chinese text-to-image model developed by China Unicom AI, and it is entirely based on domestic technology.AscendTraining and inference are implemented on an AI-based hardware and software platform. The model employs a composite language encoding module to optimize the understanding of long Chinese texts and distinctive vocabulary, thereby improving the quality of image generation. Based on massive pre-trained Chinese text and image data, Unicom Yuanjing minimizes information loss and accurately generates high-quality images. The Yuanjing text-to-image model supports domestic full-stack training and inference, adapts to custom datasets, and achieves smooth cross-platform switching. It has been applied in multiple industries such as cultural and creative industries and apparel, helping enterprises improve efficiency and reduce costs.

Main functions of Unicom Yuanjing

• Chinese semantic understandingBased on a composite language encoding module, it accurately understands long Chinese texts and words with multiple attributes, thereby improving the ability to understand Chinese semantics.
• High-quality image generationGenerates high-quality corresponding images based on Chinese text, supporting Chinese-specific vocabulary and expressions.
• Pre-training and InferenceWe introduce massive amounts of Chinese text and images to pre-train the data, thereby improving the model's ability to understand Chinese proper nouns and complex expressions.
• Computing power adaptationTraining and inference are implemented on the domestic Ascend AI basic software and hardware platform, adapting to domestic computing power.

The technical principles of Unicom Yuanjing

• Composite Language Encoding ModuleThe SDXL architecture integrates a composite language encoding module, replacing the English CLIP model with the Chinese CLIP model, thereby enhancing the ability to understand short Chinese texts.
• encoder-decoder architecture: Introduces a language model based on an encoder-decoder architecture into the language encoder part, supporting long text inputs that exceed the CLIP length limit.
• Ascend AI computing power clusterIt enables model training and inference on the Ascend AI large-scale computing cluster, providing powerful computing support.
• Interface aligned with DiffusersThe model inference interface is aligned with Diffusers, simplifying the usage process. It supports single-card and multi-card inference, and single-card inference supports UNet Cache acceleration.

Project address of China Unicom Yuanjing

• GitHub repository:https://github.com/UnicomAI/UniT2IXL
• HuggingFace model library:https://huggingface.co/UnicomAI/UniT2IXL

Application scenarios of China Unicom Yuanjing

• Digitalization of cultural relicsUsing the Unicom Yuanjing Wensheng Image Model to generate 3D images of cultural relics, providing virtual exhibition services for museums, allowing visitors to browse and learn about cultural relics online.
• Personalized clothing customizationBased on the client's specific needs, we use models to generate design sketches for customized clothing, providing clients with personalized clothing design services.
• Smart home product designGenerate concept images of smart home products based on models, helping designers quickly validate and iterate product designs.
• Advertising creative generationIt provides advertising companies with a service to quickly generate creative advertising images, automatically generating attractive visual content based on advertising copy.
• Online education platformOn online education platforms, teaching aids are automatically generated based on the teaching content, which improves students' learning interest and effectiveness.