Generating the clip was the easy part
Generating the clips took minutes. Editing took a few extra hours. Six lessons on AI video editing from a Montréal creative studio, plus what we’ll try next.
A few weeks ago we wrote that your Instagram page is not a website. This post is about the other half of our work: video.
This week we made two short videos. One was a 24-second Reel for a sports club, cut from eight phone clips and three photos. The other was a music video for an original song, with a dancer generated by AI. One of us also runs a YouTube channel, Mira Makes, where some of the same tools get tested in public.
We expected the generating to be the hard part. It wasn’t. Each generated clip came back within a few minutes. Then came a few extra hours, most of them spent cutting, syncing to the music, fixing small errors and rebuilding versions. That is the work an editor has always done, with a few new problems on top.
Cutting on the beat
The sports Reel runs on a song. Before placing any clip, we found the song’s beat grid and the moment of each racket hit in the clips’ own sound. Every cut lands on a beat, and every hit lands on a beat with a small zoom.
Choosing the part of the song mattered as much. The song is just over two minutes long. We transcribed it with a time for each word and kept about 24 seconds that end right after a full line. Then we transcribed the finished video to check that the last line wasn’t cut off.
One upload instead of eleven
The Reel used eleven short pieces from the eight clips, and we wanted them upscaled for the final edit. Rather than send eleven files, we joined the pieces into one 35-second file, wrote down the first frame of each piece, upscaled that file with Topaz at 1080p, and cut it back into eleven pieces using those frame numbers.
Before trusting the cut, we tested it on the original file. A split piece matched its source at over 50 dB on a similarity test, and a pairing we made wrong on purpose scored 17. The upscaled file came back with the same 1,050 frames we sent, so the frame numbers still lined up.
A quarter-second glitch, four attempts
In the music video there is a sunset shot of the dancer walking away. For seven frames, a little under a quarter of a second, her legs swap places.
Harout: I caught it at 1:01, and again just before 1:19, where the same shot comes back in slow motion. It’s a tiny glitch, a few hundredths of a second, but the video is better without it.
The fix took four attempts. First we played the clean first second twice. Then we froze on the last clean frame, but the freeze came out empty and pushed everything after it out of place, so we rebuilt it from a saved still. The slow-motion version still blended in bad frames, because slow motion mixes each frame with the next one, so we stopped it three frames earlier. After each attempt we checked that exact window frame by frame.
Our first layout
Our first cut of the Reel reused the layout of an earlier video: the footage inside an arch, with cards around it.
Harout: I looked at it and asked for something more creative than a design we had already used. At the least, add rackets and balls.
The second cut was built around tennis: the footage full frame, a coach-and-student split screen, photos dropped like cards on a drawn court, and a giant ball rolling across the screen between sections. We also put a scoreboard at the top that counted the sections like a match.
Harout: I wasn’t sure what the score was doing there, and sometimes it covered the players’ heads.
It became a small logo and the section name, placed higher.
Asking for a dance
For the first dance shot we wrote out the moves we wanted and gave the model (Seedance) six seconds of the chorus, which is fast. The result was fast and generic. Slowing it to 70% in the edit helped a little, and a few short pieces of it are still in the long version.
The next idea came from Harout’s review: give the model the music and let it dance to it, instead of telling it what to do. We tried that with Wan, using a slow piano section of the song and a short prompt asking her to dance freely to the music, with no list of moves. It came back smooth and graceful, and so did a ballet version made the same way. Both are in the full-length video. The short version we posted has no dance at all.
A creator we follow, AI Workflow Pro, made the same point in a recent video about editing long videos with AI: a loose brief that set only the style and the goal gave him a better edit than a detailed one. What we kept fixed was the character: one reference picture of her face for every scene.
Rebuilding any version
By the end, the music video’s edit script had a switch for each variant: the added scenes, the glitch fixes, the closing credit, the full-length song. Any version can be rebuilt from scratch in a few minutes, and we kept all of them, six of the short version and three of the long one.
Twice, a code change failed without stopping the script, and it rendered the unchanged video under a new version number. Both times the error showed in the log, and we rendered again before anyone watched it.
What we will try next
Two ideas from that same video, and one of our own. We haven’t done any of them yet.
- Have a different model watch the finished cut and describe it the way a viewer would, instead of asking the model that made it to score its own work.
- Test one short segment in the chosen style before cutting the whole video.
- Move the dancer with a real reference video of someone dancing. We tried to check the price, and the server returned an error.
We’ll write about them once we have.
Want to see what a short video could look like for your business? Message us on WhatsApp or write to studio@sociaphase.com.
Sociaphase is a creative and social media studio in Montréal. We make websites, short videos and social content for local businesses.
A note from Harout. I build software for a living, and I spent most of this week reviewing, not generating. At the end of it, I asked Claude Code to turn everything it had learned along the way into a reusable skill, so the next video starts where this one finished. AI agents didn’t replace the judgement; they gave one small team the reach to act on it. That’s what I’ll keep writing about.