In mid-October a clip went viral: a phone sits on a fridge shelf behind a tray of tiramisu, someone opens the door for "one bite" and eats the whole thing. Within days people were making AI versions with their cats.
I run a small monorepo where each trend gets its own site: shared packages for auth, a credit ledger, billing and async video jobs, and one TanStack Start app per site on Cloudflare Workers (D1 + R2). A new site is mostly a config file, a prompt, a reference clip and its own design.
This one is tiramisufridgeai.com: upload a photo of yourself or your pet and get the 10-second fridge-POV video.
Some things I learned:
- Generation is Seedance 2.5 in reference mode: your photo plus a 10 s reference clip. Reference mode copies whoever is in the reference clip, so I don't use creators' originals; the reference clips are my own smoke-test outputs with AI-made people. - Pets needed their own prompt. Without it the model gives cats human hands to hold the spoon. - The trend started from a Spanish-language reel and no competitor had a Spanish page, so the home page also exists at /es. - Money code is the boring part that matters: every credit write has an idempotency key, webhooks only trigger a re-read from Stripe, and failed jobs refund automatically from their final state.
It's paid (no free tier, since every generation costs real money), but you can watch the sample videos on the home page without signing up.
Happy to answer questions about the stack, the unit economics, or what it's like chasing trends that last a few weeks.
In the pet version, I see a Tiramisu wall in the middle of the tray!?!?
Have you tested it with different skin colors? I expect people to gt mad if the hand has the wrong color.
Is it possible to make other patterns for eating the Tiramisu? Probably too hard, and in 2 weeks the trend will be over.
Isn't that tooooooo much Tiramisu?