Tencent Hunyuan team has officially released its latest video generation model, HunyuanVideo1.5, marking another important breakthrough in video generation technology. This lightweight model, based on the Diffusion Transformer (DiT) architecture, has 8.3B parameters and can generate high-definition videos of 5 to 10 seconds. It is now available on Tencent's "Yuanbao" platform for user experience.
HunyuanVideo1.5 supports multiple generation methods. Users can generate videos from text descriptions (Prompt), or combine uploaded images with text to easily convert static images into dynamic videos. This innovative technology not only meets the needs of both Chinese and English input, but also demonstrates consistency between images and videos, ensuring that the generated videos match the original image in tone, lighting, scene, subject, and details.
In practical applications, users can generate complex scenes based on prompts. For example, a prompt describes how a miniature English garden grows inside a suitcase, and the model can accurately present this process, demonstrating a high level of instruction understanding and compliance. In addition, HunyuanVideo1.5 supports realistic and animated styles, and can generate Chinese and English text in videos, greatly enriching the possibilities of content creation.
Technically, HunyuanVideo1.5 adopts an innovative SSTA sparse attention mechanism, significantly improving inference efficiency, and combines a multi-stage progressive training strategy, achieving commercial-level performance in key dimensions such as motion coherence and semantic compliance. The deployment threshold of this model has been significantly reduced, requiring only a consumer-grade graphics card with 14G VRAM to run smoothly, allowing every developer and creator to participate in the innovation of video generation.
It is reported that previous open-source SOTA flagship models in the field of video generation usually required over 20B parameters and 50GB of GPU memory. The release of HunyuanVideo1.5 not only achieved a qualitative leap in generation effects, but also found a balance between performance and size. The model has now been uploaded to Hugging Face and GitHub, and developers are welcome to download and experience it.
With the release of HunyuanVideo1.5, Tencent further strengthens its leading position in artificial intelligence and video generation, providing content creators with more powerful tools and infinite creative possibilities. In the future, as technology continues to develop, the application scenarios of video generation will be more widespread. We look forward to HunyuanVideo1.5 bringing new changes to the industry.
