Tencent's Hybird Big Model Launches the New Era of "Wenshengtu": Generating Any Image You Want

Tencent's MixNet model recently upgraded its "WenShengTu" feature, which can generate images of various styles and textures based on user input. This feature is supported by Tencent's self-developed large-scale language model and has reached the level of GPT3.5, with a 20% improvement in code capability. The technical challenges of the "WenShengTu" feature lie in the semantic understanding of prompt words, the rationality of generated content, and the quality of generated images. To address these challenges, Tencent has proposed a series of original algorithms. In terms of semantic understanding, Tencent MixNet adopts a fine-grained model for both Chinese and English, achieving bilingual understanding and enhancing the model's perception of details and generation effects through optimized algorithms. In terms of content rationality, Tencent MixNet enhances the model's perception of two-dimensional spatial positions in images and incorporates prior information such as human skeletons and hand structures into the generation process, making the generated image structures more reasonable and reducing error rates. In terms of visual texture, Tencent MixNet improves the generation quality through a multi-model fusion approach. After optimizing the model algorithms, the portrait model of MixNet's WenShengTu has achieved a 30% improvement in details such as hair strands and wrinkles, while the scene model has achieved a 25% improvement in details such as vegetation and ripples. To generate images of different styles and textures, users can include corresponding descriptions in the prompt words, such as "realistic," "photographic," "oil painting," or "cyberpunk." Users can also provide detailed descriptions of the scene to achieve satisfactory results. Currently, Tencent has integrated MixNet into multiple development platforms internally, with over 180 internal businesses accessing MixNet, including Tencent Meeting, Tencent Docs, WeChat Work, Tencent Advertising, and WeChat Search. Some netizens have provided feedback on the internal testing and shared relevant operation pages. From the screenshots, it can be seen that MixNet supports functions such as formulating interview outlines, travel plans, PPT outlines, and fitness plans, with options for work, programming, marketing, and life. In addition, MixNet also supports AI conversation functionality, allowing users who have been approved for internal testing to experience it.