Qwen-Video-Edit turns an image editing model into a video editor

Alibaba has introduced Qwen-Video-Edit, an approach that applies the instruction-based image editing model Qwen-Image-Edit to video without training a video generation foundation from scratch. The latent frames of the Wan 2.1 video encoder are laid out as tiles of one large virtual image, and a pair of lightweight projections connects the two latent spaces. The system is fine-tuned on the public Ditto-1M dataset, and several denoising steps of Wan 2.2 serve as an optional temporal enhancer. Details are in the paper on arXiv.

What it means

The news directly concerns Wan from our reference: Qwen-Video-Edit uses the video VAE of this very model to encode frames, and Wan 2.2 is used as an optional temporal enhancement stage. Wan itself is described in our reference as a video generator with a text-to-video task, open source, and weights under Apache-2.0. Its fact sheet does not mention instruction-based editing — this is a separate capability that the new work adds on top of the existing architecture. For the AI tools for video section, this is an example of how an image editing model can be reused for video tasks without fully retraining a video diffusion foundation. The developer emphasizes that even without fine-tuning, the stock image editor edits video presented as a contact sheet, and fine-tuning closes the remaining quality gap. For now, this is research work: independent tests will show how applicable the approach is to real-world video editing scenarios.