Long-video editing — every 45-frame segment edited with a different instruction. Drag to compare.
Portrait videos supported natively (aspect-preserving).
Video generation models read and write video-VAE latents. We teach Qwen-Image-Edit's transformer to edit those latents directly: two tiny projections bridge Wan 2.1's latent space into the DiT's token space, warm-started from the DiT's own input/output layers so that a static video is embedded exactly like an image the model already understands. The latent frames are arranged as tiles of one big virtual image — the same positional treatment the image model was pretrained on. Fine-tuned with LoRA or full parameters on Ditto-1M (source, edited, instruction) triplets, then refined by a few steps of Wan 2.2 denoising-enhancement.
Drag the divider on any clip to compare the original (left) with the edited result (right).
@misc{qwenvideoedit2026,
title = {Qwen-Video-Edit: Instruction-based Video Editing by Repurposing an Image Editing Model},
author = {Yunpeng Bai, Yossi Gandelsman, Michaël Gharbi, Qixing Huang},
year = {2026},
url = {https://github.com/yunpeng1998/Qwen-Video-Edit}
}