What is Unicom Yuanjing
Unicom Yuanjing (UniT2IXL) is a Chinese native text-to-image model launched by China Unicom AI, fully implemented on domesticAscendAI basic software and hardware platforms for training and inference. The model adopts a composite language encoding module to optimize understanding of long Chinese texts and unique vocabulary, improving image generation quality. Unicom Yuanjing is based on pre-training with massive Chinese image-text data, reducing information loss and accurately generating high-quality images. The Yuanjing text-to-image model supports full-stack domestic training and inference, adapts to custom datasets, and enables smooth cross-platform switching. It has been applied in multiple industries such as cultural creativity and clothing, helping enterprises improve efficiency and reduce costs.
Main Features of Unicom Yuanjing
- Chinese Semantic Understanding: Based on a composite language encoding module, it accurately understands long Chinese texts and multi-attribute corresponding vocabulary, enhancing Chinese semantic understanding capability.
- High-Quality Image Generation: Generates high-quality corresponding images based on Chinese text, supporting Chinese-specific vocabulary and expressions.
- Pre-training and Inference: Introduces massive Chinese image-text pair data for pre-training, improving the model's understanding of Chinese-specific nouns and complex expressions.
- Computing Power Adaptation: Implements training and inference on domestic Ascend AI basic software and hardware platforms, adapting to domestic computing power.
Technical Principles of Unicom Yuanjing
- Composite Language Encoding Module: Integrates a composite language encoding module into the SDXL architecture, replacing the English CLIP model with Chinese CLIP to enhance understanding of short Chinese texts.
- encoder-decoder Architecture: Introduces a language model based on the encoder-decoder architecture into the language encoder part, supporting long text input exceeding the CLIP length limit.
- Ascend AI Computing Cluster: Implements model training and inference on Ascend AI large-scale computing clusters, providing powerful computing support.
- Interface Aligned with Diffusers: The model inference interface is aligned with Diffusers, simplifying the usage process, supporting single-card and multi-card inference, and single-card inference supports UNet Cache acceleration.
Project Address of Unicom Yuanjing
- GitHub Repository：https://github.com/UnicomAI/UniT2IXL
- HuggingFace Model Library：https://huggingface.co/UnicomAI/UniT2IXL
Application Scenarios of Unicom Yuanjing
- Cultural Relic Digitization: Use the Unicom Yuanjing text-to-image model to generate 3D images of cultural relics, providing virtual exhibition services for museums and allowing audiences to browse and learn about cultural relics online.
- Personalized Clothing Customization: Based on customers' specific needs, use the model to generate design drawings for customized clothing, providing personalized clothing design services for customers.
- Smart Home Product Design: Based on the model, generate concept images of smart home products, helping designers quickly validate and iterate product designs.
- Advertising Creative Generation: Provides advertising agencies with services to quickly generate advertising creative images, automatically generating attractive visual content based on advertising copy.
- Online Education Platforms: On online education platforms, automatically generate teaching auxiliary images based on teaching content, improving students' learning interest and effectiveness.