We Spent $180 Teaching an AI to Direct Music Videos. Here's What Makes Them Look Like AI Slop.
October 5, 2026 · 4 min read · 10 versions of the director, 45+ graded drafts
Same song. Left: our AI director before the test. Right: the same director after it.
Same song, “Calm in the Center”. Both untouched, straight out of the director.
Try the new director on your song ↓The left one looks expensive. It also makes no sense. Helmets, sparks, lava, six different people — for a song about breathing slowly.
So we spent two days and about $180 in rendering fixing it: real songs, every draft graded by a person watching on a phone. Here's what we learned.
The short version
Too many people = instant AI slop. Random humans in random clothes were the #1 complaint. Show the figure rarely, and make it count.
One object beats ten pretty shots. A lantern that travels through the whole video feels directed. Ten unrelated beautiful shots feel generated.
A single physics mistake kills the video. A truck rolling backwards turned a 9/10 into a disaster. People forgive style, not physics.
Pretty is not the goal. Fitting the song is. Every good video in the test came from reading the lyrics first and choosing one idea — not from a better prompt.
Want to see what the director makes for your song?
Upload a track. Your analysis and 7 short clips are free. The AI video is one click on your results.
AI video under $10 · no subscription · refunded if it fails
The 7 rules we ended up with
Each one exists because a real video broke it. They're now built into the director, so nobody has to remember them.
“Too many people” was the most common complaint. Ordinary people in everyday clothes look random. Now the main figure is in at most 2 of 8 shots, and it's someone with presence: a stone soldier, a ghost, a lighthouse keeper.
A masked or hooded figure looks mysterious instead of fake. That one rule fixed most of our rap and metal videos. It never applies to sad songs — there it turns creepy.
A lantern, a red thread, a spear. When one thing travels through every scene, the video feels directed. When ten beautiful things appear once each, it feels generated.
Random walking is AI video's filler. Every figure now does one thing that means something, and the camera is never in the shot.
A trap song about setting trends came back as a matador. Grade: 1/10. Since then, nothing gets a costume the lyrics don't ask for.
A truck rolling backwards or boots on the wrong feet ruins a video in one second. Viewers forgive style. They don't forgive physics.
No shot longer than about 4 seconds. Lyrics big and bold, the sung word lit in a colour taken from the video itself — fixed orange looked cheap on every song.
Did it work?
Mostly. Grades went from 5.0 to 7.0 — and the final round used songs the director had never seen.
Average grade, out of 10
The number that matters is the last one: 7.0 on songs the director had never seen, so we weren't just tuning it to our test set. (The full set averaged 6.1 before the fix.) Best single video: 10/10 (a song about glass wings, spirits instead of people).
Three videos from the test
All our own songs, all one click, nothing edited afterwards. Two hits, one miss.
We kept Old Truck Home in on purpose. Its stills scored 9/10; the video was a disaster. That's why every vehicle now only drives forwards.
The surprise: the AI critics were wrong
We built automatic critics to grade the drafts. They're great at spotting defects — a sixth finger, text in the image, a pose. But on “is this a good video for this song?” they disagreed with a human watching on a phone again and again.
So the critics now check every image before it gets animated, and the taste rules come from people. If you're making AI videos yourself: trust your thumb on the scroll, not a score.
$180 of mistakes, one button for you
Upload your song, press “AI video”, and about 8 minutes later you have a finished vertical video with lyrics.
Make My AI Music Video →Common questions
Why do AI music videos look like AI slop?+
In our test it came down to three things: too many random people on screen, shots that have nothing to do with the song, and physics mistakes like a car rolling backwards. Style was almost never the problem.
How do you make an AI music video look good?+
Start from the song, not from a prompt. Pick one idea for the whole video and one object that carries it, show the main figure rarely, never let anyone pose, and check every shot for physics before you animate it.
How much does an AI music video cost to make?+
Our two days of testing cost about $180 in AI rendering. A finished video on Songbrain costs you under $10 — the $10 credit pack covers it — and the credits come back if the render fails.
Did an AI judge the videos?+
Partly. Automatic critics checked every still for defects, but the final grades came from a person watching on a phone — the critics disagreed with human taste too often to trust them alone.