Qwen-Video-Edit: Instruction-Based Video Editing for Wan Latents

Variants

Select a variant above for curated weight downloads, tutorials, and related information.

About Qwen-Video-Edit

Qwen-Video-Edit takes a different route than video-pretrained editing models. Instead of adapting a video diffusion transformer to follow editing instructions, it teaches a strong instruction-based image editor to edit video latents: Qwen-Image-Edit's DiT operates directly on Wan 2.1 video-VAE latents, bridged by a pair of lightweight input and output projections that are warm-started from the editor's own patchify and unpatchify layers. At initialization a static video is embedded exactly like an image the editor already understands, so fine-tuning only has to learn the distribution shift.

Three design choices carry the method. The latent frames are arranged as tiles of one virtual large image, so the editor's image positional encoding applies to every tile unchanged. A Qwen2.5-VL branch encodes the instruction together with a preview grid of the source frames. The frozen Wan 2.1 VAE handles encoding and decoding on both ends, and an optional Wan 2.2 denoising-enhancement stage taken from Ditto cleans up the edited latents. The report is by Yunpeng Bai, Yossi Gandelsman, Michaël Gharbi and Qixing Huang, and the code, paper and 360P checkpoint landed on 2026-08-14, with ComfyUI support and the 480P and 720P checkpoints following on 2026-08-20.

Checkpoints

Fine-tuned checkpoints are published from 360P to 720P. Directory names encode the Ditto-1M training subset (global, local, sim2real) and the supported chunk length (45 or 81 frames), and the checkpoint trained on the global plus local subsets at 360P is the recommended default. At inference the frame count and the maximum pixel budget must match the checkpoint, and every checkpoint uses latent_mode wan_compressed with pe_mode grid.

ComfyUI

The repository doubles as a ComfyUI custom-node pack with single-chunk semantics: one sampler call edits one num_frames window, and long videos are edited chunk by chunk. Three nodes are registered under the QwenVideoEdit category, and the pack depends on ComfyUI-VideoHelperSuite for video loading and combining. The base Qwen-Image-Edit and Wan 2.1 weights are fetched automatically on the first run, roughly 55 GB in total, and the enhancement stage additionally needs Wan 2.2 T2V-A14B. Recommended checkpoints are also mirrored on ModelScope.