A clay kiwi, a robot in a tomato hat, and why take one won

I made a sleepy kiwi, a proud robot and a cheeky cat out of clay last month, and I never touched any clay. Not a crumb of it. The fingerprints you can see on them are fake too, which is a strange sentence to type on a Sunday. I also made a short video of this, episode one of a Sunday series I call Creative AI Tests. This is the written version, with the bits that didn't fit. Three scenes, all in a handmade stop-motion style. A kiwi with its beak in a coffee mug, just before the toaster goes off. A robot holding up its very first sandwich, proud as anything, with the tomato about to slide out. And a cat nose to nose with a ladybird on a laptop. A bug in the code, basically. It took two steps. Step one is the picture. You type a short line, something like "a clay kiwi asleep in its coffee". The model has learnt from loads of captioned images what clay and coffee mugs look like, so it makes a brand new picture to match. Fair warning: the first picture is rarely the keeper. I made a few for each scene and picked the best. Step two is the video. You hand the model that picture with another short line, like "the toast pops and the kiwi wakes up". The picture becomes frame one, so the kiwi stays the same kiwi, and the model works out the movement. Even the sound, and a line of dialogue. The kiwi mumbles "Five more minutes." The robot announces "Nailed it!" The cat goes with "Not today" and then, after things go sideways, "I meant to do that." (I'm not sharing my full prompts. They're less magic than you'd hope, and I like keeping a bit of mystery.)

Right from take one, the look was there. Fingerprints in the clay, little felt and cardboard sets, soft shadows, that slightly jerky stop-motion movement. All three scenes look like they came off the same table, under the same lights, and I didn't have to fight for any of it. Most of the gags landed too. The toast popped. The ladybird landed on the cat's head. And the robot ended up wearing a tomato slice as a hat, which was never in the plan. I kept it. Some accidents deserve a pay rise. Then the work started. I'd asked for tall, vertical pictures. All three came back wide, so each one had to be cropped before it could become a video. Not a big job, just the first sign that the generator and I had different ideas about the word "tall". Next came the voices. I wanted each character to sound like itself, so I ran all three scenes again. Take two got the voices right and broke the video. The robot's sandwich simply vanished in mid-air, the way lunch does in an office fridge. So I watched every clip closely and wrote a fix list. Odd artefacts. A yawn. A toast pop that didn't look right. A sandwich that wouldn't co-operate. A cat whose mouth didn't match its words. Take three went in carrying all of that, and came out worse. More artefacts, a weaker toaster sound, one slice of toast instead of two, and the cat and the ladybird had lost whatever it was they'd had. Even the cuts were off. Mr. Mummu, who wanders into my stories now and then, has the same problem with voice notes. He once recorded a birthday message for an aunty in a group chat. Take one was warm and a bit croaky. Take two had a dog barking. By take five he was so careful he sounded like a bank's hold music, so he kept going. At take nine his partner leant over and said, "Just send the first one." Mummu scrolled up. He had sent the first one forty minutes earlier, by accident. The aunty had already replied: "Lovely, so natural." He put the phone down and said, "So that is that."

That's the trap with fix lists. Each take chases the last flaw, and the thing that made the first one alive quietly walks out the back. In the end, take one went in for all three scenes. The gags landed best, and I liked its look and timing, even though the voices don't really suit the animals. The polish happened outside the generator. I sharpened the clips with a free tool and put the captions in myself, in the edit. Control is the honest limitation. I could describe a mood and get it back looking lovely. I couldn't reliably get two slices of toast, or a sandwich that stays in a robot's hand. Asking again didn't fix that. Sometimes it made it worse, and I had nine clips to prove it. If you need something exact, plan for a human doing the fiddly end of the job, and more patience than you think. Which is really the whole test. You get a style fast. Timing and polish still need a person, and more takes aren't always better takes. Sometimes the job is knowing when to stop. The robot still ends up wearing a tomato as a hat. I'm counting that as a win.