One Cut - Edit Video by Editing Code

One Cut will transcribe your video using fast-whisper and generate Python code like example.py. The code contains all the subtitles from ASR and allows you to modify the video by changing the code.

Run onecut to transcribe your video:

uvx onecut path/to/your/video.mp4

The OneCut CLI will generate code like this. You can easily fix ASR errors by changing the text.

movie = Movie("a.txt", [])
SUB_1 = sub("Sub1", 0, 1)
SUB_2 = sub("Sub2", 1, 2)
SUB_3 = sub("Sub3", 2, 3)

FINAL_CLIPS = [SUB_1.to(SUB_3)]

You can cut one section of your video (like SUB_2) by doing this.

FINAL_CLIPS = [SUB_1, SUB_3]

Or use the helper function to cut it, the function is super useful if you only want to remove a small portion of clips.

FINAL_CLIPS = remove_subs(movie.all_subtitles, [SUB_2])

After finishing cutting & fixing, just run python {generated_file}.py to generate the final video and subtitle(.srt) file.

Dependencies

1. Python 3
2. FFmpeg

The generated file is self-contained. It does not depend on the onecut package. The only dependency is ffmpeg.

Why?

I want to share some of my videos but the only computer I have is a Steamdeak. Running any modern video editors will kill my Steamdeak and it's super laggy as well. There are some online services like the Descript but I don't want to upload the video over my slow internet and pay $12 per month.

Also, the coding agent is very good at spotting ASR errors and I want to generate something for them to work on. All the existing tools are based on text but I think using code as intermediate representation is much better for coding agents.

Agent Edit

Most of the time, I just ask the agent to fix the ASR errors in generated.py and it will think for a long time and fix most of the ASR issues. You can ask the agent to do additional things like:

1. Burn-in the subtitles.
2. Remove part of video.
3. Only export part of the video.

Basically, the agents can do anything supported by the FFmpeg filter.

Using Deepseek-V4 Flash, fixing all the ASR errors costs about $0.1 in API pricing.

Future Plans

1. Support Word Level Editing: I don't know what's the best way to support that since I don't want to put a lot of timestamps in the code.
2. Bugfix

LLM Usage

This project is written in the old-fashioned way. LLM is used for data collection and some debugging.