About Kiwi-Edit
An open-source framework for editing video with plain language — and with reference images when language isn't precise enough.
What Kiwi-Edit is
Kiwi-Edit is a video editing framework published by Show Lab. It couples a multimodal large language model encoder with a video diffusion transformer: the encoder interprets your instruction against the source footage, and the transformer regenerates the clip with that change applied.
It supports two ways of asking for an edit. Instruction editing works from text alone — restyle a clip, add or remove an object, replace a subject, change a background. Reference-guided editing adds an image, which is what you need when the target is a specific person, product or location rather than a generic one.
How it works
The instruction, the source video and any reference image go through a Qwen2.5-VL-3B encoder, which resolves what you are pointing at and what you want done to it. A Wan2.2-TI2V-5B diffusion transformer then generates the output, conditioned on the original clip so that motion, framing and the regions you did not mention carry through unchanged.
Training happens in three stages — images first, then images with video, then reference-video data — at up to 720×1280 resolution and 81 frames. The published checkpoints reflect those stages: one fine-tuned on instruction data, one on reference data, and one on both.
How it is evaluated
Two public benchmarks cover the two modes. OpenVE-Bench measures instruction-based edits; RefVIE-Bench pairs source videos with reference images to test whether the specified subject or background actually appears in the result.
About this site
kiwiedit.com is an independent informational resource. It explains what Kiwi-Edit does, how to prompt it, which checkpoint to pick and how to run it locally, and it embeds the authors’ official Hugging Face Space so you can try the model without installing anything.
This site is not affiliated with, endorsed by or operated by Show Lab, Hugging Face or the Kiwi-Edit authors. It does not host the model, process your videos or store your uploads — any edit you run goes directly to the Hugging Face Space, under Hugging Face’s own terms. For authoritative details, licences and updates, always refer to the official sources.