Turbo Was the Slow Lane

One man, one park bench, ten thousand years, and one evening at the computer

The video that started it

Maarten Albers shared a one-minute clip for the weekly challenge. A bearded man sits on a bench and never moves. The camera never moves either. Everything around him does. His furs melt into an Egyptian tunic, the pyramids fold themselves into the Colosseum, a medieval market rises out of the dust, then Victorian London in the rain, a 1960s park with a camper van, a city of glass towers, a neon night, and finally a gleaming city of gardens and waterfalls. No cuts. History just flows past him like water around a stone.

Maarten explained that the effect came from a LoRA called Animatediff style for MiniMax H3. Think of a LoRA as a small add-on that teaches a big AI video model one particular style, the way a filter changes a photo. This one was made for MiniMax H3, a video model I can run on my own computer for free. Its speciality: scenes that keep reshaping, like that clip.

I found the effect so cool that I decided to reverse engineer it (work out, step by step, how it was made) using Claude AI, the AI assistant I use in my own work.

The first try, and what went wrong

AI video works in short pieces, about five seconds at a time. You give it a starting picture and a description of what should happen, and it renders (computes) the clip frame by frame. My first five seconds, Stone Age to Egypt, actually worked. The fire, the hide tents and the cliff melted into sandstone, palm trees and pyramids, and a glowing blue orb formed in his hands.

But two things were off. By the end, his short curly hair had grown long and straight. A different man. And the camera had slowly crept closer, so his feet were touching the bottom of the frame. Each new clip starts from the last frame of the one before, so every small mistake gets passed down the chain and grows. After ten eras he could have been a stranger in a close-up.

The shortcut

The clip took about four minutes to make. I had nine more to go, so I asked for the "turbo" setting, an add-on that promises the same clip in far fewer steps.

It did the opposite. A normal step took about 8 seconds. With turbo switched on, the first step took 81 seconds and the second took about 170. The graphics card was showing 100% busy while drawing barely a third of its normal power. It wasn't working; it was shuffling memory in and out, like a desk too small for the papers on it. The clip would have taken a quarter of an hour. We cancelled it. Notes from earlier projects had warned about exactly this, and they turned out to be right.

The fix: pictures you can pin down

What actually solved the problem was going slower and planning more. I switched my picture-making to Nano Banana 2, the image tool I use for my music videos, and had it paint ten key frames: the same man, on the same spot, in the same pose, once for each era. Rome with a brass armillary sphere. Medieval with a pewter tankard. A ruff collar and a globe in a Dutch harbour. A letter under a London gas lamp. Tie-dye and a guitar. A phone. A VR headset. Neon glasses in the rain. White robes beside a waterfall. Those ten pictures cost about 300 credits on OpenArt.

Then each clip was told where to start and where to finish. The video model fills in the morph, but it has to land on that exact picture, so the man and the framing get reset every five seconds. The drift can't build up.

A small script then rendered the remaining eight clips one after another, about four minutes each, while I was away from the desk. Another one assembled everything into a CapCut project (CapCut is the video editor I finish everything in), with my logo in the corner and the title "Man On Parkbench".

How long it took

Every file the computer saved carries a time stamp, so here is the evening, to the minute:

• 9:19 pm Maarten's video lands in my Downloads folder.
• 9:24 pm Claude has pulled frames out of it and worked out what happens, era by era.
• 9:28 pm The LoRA is automatically downloaded and installed.
• 9:31 pm The first starting picture of the Stone Age man is ready.
• 9:37 pm The first test clip finishes: Stone Age to Egypt, with the drifting face and the creeping camera.
• Around 9:40 pm The "turbo" experiment, abandoned within minutes.
• 9:43 pm The new Stone Age picture from Nano Banana 2.
• 9:48 pm The Rome key frame.
• 9:53 pm Clip 1 re-made from the new picture.
• 9:56 to 9:58 pm The other eight key frames, Medieval to Utopia.
• 9:58 pm Clip 2 proves the pinning trick works: Egypt melts into Rome with no cut.
• 9:59 pm The script starts rendering clips 3 to 10 on its own.
• 10:19 to 10:39 pm While it's still running, I upscale each finished clip in Magnific (an online tool that redraws video at a higher resolution) as it comes out.
• 10:30 pm Clip 10 finishes. One second later, the CapCut project is built, branded and checked.
• 10:43 pm The sharper Magnific clips are swapped into the CapCut project.
• 10:46 pm I watch it through in CapCut and export the finished film at 1440p.

From download to finished film: 1 hour 27 minutes. About 40 minutes of that was the computer rendering, and for half an hour of it nobody had to touch anything.

Where it stands

Ten clips, about 51 seconds, no cuts anywhere. He's recognisably the same man from the Stone Age to the far future.

It's exported, enhanced and sharpened with Magnific to 2K aka 1440p, and uploaded to YouTube. It isn't flawless: the joins between clips are close but not pixel-perfect, so at one or two of them you may catch a split-second hiccup. Watch for them.

The takeaway

The speed setting was the slow road, and the slow road (painting every destination before setting off) was the fast one. When a tool promises a shortcut, check where it actually goes before you take it.