Qwen-Video-Edit repurposes an image editing DiT to edit videos, no video-pretrained backbone needed
qixing_huang · x · 2026-08-18
Qwen-Video-Edit introduces instruction-based video editing by repurposing the pretrained image editing model Qwen-Image-Edit's DiT to operate directly on Wan 2.1 video-VAE latents — no video-pretrained transformer required.
- Method: 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 a static video embeds exactly like an image; latent frames are laid out as tiles of one big virtual image with grid RoPE, matching the image model's pretraining. Prompts come via Qwen2.5-VL.
- Training: 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.
- Capabilities: long-video editing with a different instruction per 45-frame segment, and native aspect-preserving portrait video support.
Code and the Hugging Face model are released.
More from Multimodal
- FLUX 3 Upscaler hits Higgsfield: 50% faster and cheaper, native 2K/4K video — tisch_eins · 2026-10-16
- AI-generated "RiRi" images stun with photorealistic diamond-gown Rihanna looks — egeberkina · 2026-10-16
- SeedAudio 1.5 hits CapCut PC: one prompt generates dialogue, SFX and music — LearnWithBishal · 2026-10-16
- DLSS 5 Assetto Corsa demo stuns sim racers: 'better than reality' — CSProfKGD · 2026-10-16
- Video recast with fal agent: swap a character via one character sheet — adamho · 2026-10-16
- Dev open-sources Krea 2 pixel-art LoRA for 16x16 game icons with transparent backgrounds — Angaisb_ · 2026-10-16