Lev
Gartman

I love telling stories about dogs, and I love working AI into production.

The process followed a traditional shape: script, then a character and environment sheet, then an animatic storyboard, then animation. About sixteen working days of actual production from blank page to finished short.

The hard parts

The shots that fought me hardest were the ones that asked the AI to reveal something the starting frame didn't already contain. A character starting with their eyes closed and then opening them (waking up), when the eyes weren't yet drawn. Walking through saloon doors into a saloon interior that didn't exist. Some were a challenge even when using an ending frame. The fix was sometimes more text description, and sometimes less. And sometimes just a lot of re-rolling.

Another fix when the physics of a shot were particularly difficult: generate the reverse action, then play it backward in Premiere.

What stood out

One delight: generating with a starting frame but no ending frame, and watching how the AI chose to "perform" the actor or move the camera. Plenty of misfires, but every so often a happy accident I'd never have thought to direct.

Another thing that surprised me was how much further these tools have come in a year. The last short I made in 2025 took noticeably more work for similar results. The capability curve is real and steep.

What hasn't changed: the things that make a short worth watching. A keen eye. A sense of humor. Twenty years of video production and editing telling you when a shot is working and when it isn't. The tools are catching up fast. The craft still has to lead.

"Dogtown just ain't for suckers."

The whole thing ran on a shot-by-shot storyboard document, twelve shots, each one carrying its own first-frame image prompt and its own video prompt as separate columns. That separation is the part that matters. Directing a generative shot is two different jobs: establishing what the frame contains, then establishing what moves. Collapsing them into one prompt is where most of these attempts fall apart.

The tool I built to do it

Writing those prompts by hand for every shot, and keeping a character consistent across all of them, is the actual bottleneck. So I built a storyboard generator: paste in a multi-line script, pick a visual style, and it generates the full board as a grid, holding a reference set so the same character survives from shot to shot.

It was meant to run on Adobe Firefly's API. Firefly turned out to have no self-serve developer access at all, only an enterprise agreement starting around a thousand dollars a month with a multi-week sales cycle. So I rewrote it against Google's Gemini image model, which gives you a single API key and a billing toggle. That detour is worth stating plainly: the constraint that decided the architecture was commercial, not technical.

What it cost

Eight working days spread across a month, with the real sprint in the final week and the export finishing at 11:22pm on deadline night. That is the honest number, and it is half what the claymation short took, because the second time through you already know which shots are going to fight you.