wan-video/wan2.1

Wan2.1 is a suite of open-source models for generating video from prompts, described in the README as a comprehensive and open set of video foundation models. The repository ships the inference code and weights so that anyone can run the models locally rather than depending on a paid service. The suite covers several related generation tasks. Text-to-Video takes a written prompt and produces a clip, Image-to-Video animates a still picture, First-Last-Frame-to-Video fills in the motion between two given frames, Text-to-Image generates stills, and Video-to-Audio creates sound to match a clip. There is also a video editing pipeline called VACE, introduced as an all-in-one model for video creation and editing. Underneath these tasks sits Wan-VAE, a video encoder-decoder that can compress and reconstruct 1080P videos of any length while keeping their temporal information intact, which is what lets the higher-level models work efficiently. One advertised feature is that Wan2.1 can render readable Chinese and English text inside generated video. A smaller 1.3B-parameter variant of the text-to-video model is sized to fit on consumer GPUs, needing about 8.19 GB of VRAM and producing a five-second 480P clip on an RTX 4090 in roughly four minutes. Someone would reach for Wan2.1 to prototype short video clips from a text description, animate marketing or research stills, build tools on top of a strong open video backbone, or compare against closed commercial video generators. The code is Python and weights are also published on Hugging Face and ModelScope, the models are integrated into Diffusers and ComfyUI. The full README is longer than what was provided.