Qwen Video Edit

Qwen-Video-Edit is a research project that performs instruction-based video editing by adapting an image editing model, Qwen-Image-Edit, to work on video latents. Two small projection layers map Wan 2.1's video-VAE latent space into the image model's DiT token space, initialised from the DiT's own input/output layers so a static video is embedded like an image the model already understands. Latent frames are arranged as tiles of one large virtual image, reusing the positional treatment from image pretraining.

The model is fine-tuned with LoRA or full parameters on Ditto-1M (source, edited, instruction) triplets, then refined with a few Wan 2.2 denoising-enhancement steps. The project page shows long-video editing, where each 45-frame segment gets a different instruction, and native support for portrait (aspect-preserving) videos. Comparison sliders show original versus edited clips.

According to a YouTube creator review, code and a training script have been released, with a total download of about 41 GB. The project page lists authors Yunpeng Bai, Yossi Gandelsman, Michaël Gharbi and Qixing Huang. The provided material does not give a repository URL, license, weights location or paper link, so those details are unconfirmed.