The overall Chinese ability exceeds GPT3.5, and the code ability is increased by 20%. And officially opened the “Wensheng Map” function to the outside world.
For the new “Wensheng Diagram” function. Tencent said it had conducted special technical research and proposed a series of original algorithms:
🔍 In terms of semantic understanding
Tencent Hunyuan adopts a bilingual and fine-grained model that models both Chinese and English to achieve bilingual understanding, rather than through translation. This improves the model’s perception and generation effect of details through optimization algorithms, effectively avoiding misunderstandings under multicultural differences.
🎨 In terms of content rationality
AI-generated human structures and hands are often prone to deformation. In order to solve this problem, the Hunyuan Wensheng graph enhances the two-dimensional spatial position perception of the algorithm model and introduces prior information such as human skeleton and human hand structure into the generation process. This makes the generated image structure more rational while also reducing error rates.
🖼 In terms of picture texture
The hybrid element Wensheng graph adopts the method based on multi-model fusion to improve the generated texture. After the optimization of the model algorithm, the effect of the portrait model in terms of hair, wrinkles and other details has been increased by 30%, while the effect of the scene model in terms of grass, trees, ripples and other details has been increased by 25%.
At present, more than 180 Tencent internal businesses have been connected to Tencent Hunyuan, including Tencent Meetings, Tencent Docs, Enterprise WeChat, Tencent Advertising, and WeChat Search. Recently, QQ Browser also launched a “PDF reading assistant” based on Tencent Hunyuan, with functions such as intelligent summarization, intelligent Q&A and multiple rounds of questions.
