Alibaba Group (BABA, Financial) has announced the open-sourcing of its video generation model, Wanxiang 2.1, under the Apache 2.0 license. This release includes full inference code and weights for models with 14 billion and 1.3 billion parameters. The model supports both text-to-video and image-to-video tasks.
The 14B Wanxiang model excels in instruction adherence, complex motion generation, physical modeling, and text-to-video generation. It has significantly outperformed other models like Sora, Luma, and Pika on the authoritative VBench evaluation set, achieving an overall score of 86.22% and securing the top position.
The 1.3B version also showcases impressive results, outperforming larger open-source models and even coming close to some closed-source models. Remarkably, it can operate on consumer-grade graphics cards, requiring just 8.2GB of video memory to generate high-quality videos, making it suitable for secondary model development and academic research.
