Showlab's Kiwi-Edit brings prompt-based video editing to open source
Showlab just open-sourced Kiwi-Edit, a video editing framework that pairs a multimodal LLM with a 5B diffusion transformer to edit clips using plain text or reference images. It handles 720p style transfers, background swaps, and adding or removing objects while keeping temporal consistency. With code, weights, and datasets all released, open-source video manipulation just took a serious step forward.
Kiwi-Edit is a new open-source video editing framework from Showlab that lets you manipulate existing 720p clips using plain text instructions or a reference image.
Unlike traditional text-to-video generators that build scenes from scratch, Kiwi-Edit tackles the messier job of video-to-video editing. Best of all, Showlab fully open-sourced the code, model checkpoints, and training datasets.
What it does
Kiwi-Edit handles two main editing workflows: instruction-guided edits and reference-guided modifications.
With text prompts, you can apply global style transfers like turning real footage into a watercolor painting or an anime sketch. You can also target local details, such as putting a hat on a character or cleanly erasing an unwanted person from the background.
When you supply a reference photo, the system can extract the subject or background from the still image and swap it into the video. The model keeps the camera movement and original physics intact while preserving temporal consistency from frame to frame.
Under the hood and hardware specs
The architecture pairs a multimodal LLM for instruction parsing with a 5B parameter video diffusion transformer (DiT) for generation.
Showlab trained Kiwi-Edit through a three-stage curriculum spanning representation alignment, instruction fine-tuning, and reference-guided fine-tuning. They also built an automated pipeline to turn instruction triplets into reference training quadruplets.
On public benchmarks running at 1280×704 resolution, Kiwi-Edit outperforms several 5B and 14B open-source baselines including Lucy-Edit and OpenVE-Edit. It even challenges commercial proprietary models like Runway Aleph on specific local editing metrics, though closed-source heavyweights like Kling-O1 still hold an edge in subject identity preservation.
Running it locally requires serious muscle. The model takes roughly 20 GB of disk space and needs a high-end consumer GPU to run smoothly.
The paper is up on arXiv, and the code, weights, and datasets are hosted on GitHub and the project page.
Why it matters
Precision editing has always been the missing piece in open-source generative video. While commercial APIs offer polished tools behind paywalls, Kiwi-Edit gives developers and creators a transparent, fully reproducible pipeline that actually works on 720p footage.
We think releasing the datasets alongside the 5B weights was the right move, as it lets the community build custom video editing workflows without relying on black-box cloud APIs. The main barrier remains the hardware footprint, but for studios with serious GPUs, this is one of the strongest open video manipulation toolkits available today.
Sources (1)
- showlab.github.io(opens in a new tab)Originalshowlab.github.io