Alibaba Wan2.1: fastest and best video AI on your own PC [15:49 Tue,4.March 2025 by Thomas Richter] The Chinese internet giant Alibaba has released a new video AI that could repeat the success of DeepSeek in the LLM sector. Alibaba has released its Wan2.1 model as open source, meaning it is freely accessible and can be used and further optimized free of charge.
The highlight: The quality of the videos generated by Wa2.1 is extremely good and also very fast. According to Alibaba, Wan2.1 can generate more realistic videos, even with difficult motifs such as complex body movements, rotations, dynamic scene transitions, and smooth camera movements. Wan2.1 is also said to have an understanding of real-world physics (such as fluid dynamics) and realistic object interactions in order to simulate realistic movements. Important for good images is, of course, the ability to depict cinematic effects, i.e., special lighting situations, image styles, and camera movements. Optimized Version: 8.19 GB VRAM Sufficient And the best part: just a few days after the launch, there is already an optimized version called Wan2.1 GP , which only requires 8.19 GB of VRAM and thus runs on many consumer graphics cards, such as Nvidia&s RTX graphics cards from the 5060/4060 TI onwards. On an Nvidia RTX 4090, generating a 5-second video with a resolution of 480p takes approximately 4 minutes (without optimization techniques such as quantization). With the latest version, even 720p clips longer than 10 seconds can be generated on an RTX 4090 with 24 GB VRAM or 10-second 480p videos on a GPU with less than 12 GB VRAM. Wan2.1 includes not only a text-to-video but also an image-to-video model with 14 billion parameters (14B), both with 480p and 720p resolution. Another text-to-video model with 1.3 billion parameters (1.3B) is so small that it also runs on consumer GPUs with at least 8.19 VRAM. Object-Based Video Editing Wan2.1 can also reliably render text in videos and also includes a video-to-audio model that automatically provides suitable sound accompaniment for a video clip - a capability that only recently (until recently) has been mastered by other large video AIs like Lumas Dream Machine. cdn.wanxai.com/static/demo/output/edit-3.mp4 Automatic SFX Wan2.1 can generate both sound effects and musical soundtracks - thematically and rhythmically appropriate. Wan2.1 can also be used to edit videos via AI using image or video templates or to specifically change them via in- or outpainting. The model also seems to be more lenient regarding nude skin than other models. The Wan2.1 model, including weights, can be downloaded via Huggingface or GitHub and is also supported by ComfyUI . Better than Sora, Runway, and Luma? Wan2.1 also performs very well in a quality comparison with other video AIs such as Sora, Runway Gen 3, MiniMax, Luma Dream Machine, and Pika and achieves first place in the VBench , a benchmark for video AIs that considers factors such as image quality, consistency, human movements, and image aesthetics, although the current top performer, Google Veo 2, is not included.
VBench more infos at bei wanxai.com deutsche Version dieser Seite: Wan2.1 - beste kostenlose Video-KI für den Home-PC?
[nach oben ]