Hard Eight
Generative study 01
A generative dice roll that lands on the hard 8: a study in where AI video still breaks
- Role
- Creative / Generative / Edit
- Stills
- Nano Banana Pro · GPT Image 2
- Motion
- Seedance 2.0 · Veo 3.1
- Finished in
- Premiere Pro · Topaz Video AI
- Year
- 2026
I set out to generate it as one continuous shot.
A moody, Wong Kar-wai-style dice roll across a craps table that resolves on the hard eight: both dice landing on four. What follows is what I learned making it: where these tools help, where they break, and which problems belong to the prompt versus the edit. Two things broke most often: one I could prompt around, and one I couldn’t fully beat.
The dice physics. They wanted to slide flat, like pucks with pips. Weight was the hardest thing to get, and it came from language, not settings: describing the bounce, and killing the word “roll.”
Flat dice, no weight: the most common failure, and the hardest to prompt out.
The pips. This is the one I couldn’t fully beat. On the fast spins they miscount, and if you look closely a few wrong ones made the final cut. Knowing where the tools still can’t be trusted turned out to be as useful as knowing where they can.
Miscounted pips in the air. I caught the worst; some made the cut. Not solved.
First and last frames. I’d generate the two endpoints of a beat as stills, then hand them to the video model to travel between: my first run at the one continuous shot. When continuous didn’t hold, the same trick became how I built the shorter clips I could cut together.
The two endpoint frames of a beat: the model generates the motion between them.
That’s the whole job, really: direct the tool, judge the output, and stay honest about what it can’t do yet.
All original IP · Made under these principles