I built an AI video maker for one meme: First Time in a Fancy Cafe

Disclosure: Fancy Cafe AI is my product. This post is about how and why I built it, including the parts that might put you off.

There is a meme that took off on TikTok in early November 2026 that I can't stop thinking about. One fixed camera. A minimalist concrete cafe. A low bench, and a tiny round table mounted on the wall that looks a lot like a stool.

Someone is sitting proudly on that tiny table, feet on the bench, sipping a latte like they own the place. A stranger walks in, sits on the bench and puts their cup down on the table next to them. The first person freezes mid-sip. The cup. The table. The realisation. Then they slide down quietly, sit properly, and stare ahead as if nothing happened.

Why a whole tool for one joke

I like building small, focused tools, and the pattern I keep coming back to is: notice a trend or an everyday annoyance, build the simplest thing that solves it, and be honest about privacy and pricing.

The annoyance here is obvious once you try to join the trend yourself. You need a second person, a cafe with that kind of wall table, and someone to film it from one fixed angle. Most people have none of those. Plenty of people search for a CapCut template, but templates paste a face onto fixed frames, and this joke lives in posture and timing: the smugness on the table, the stiff freeze, the slow slide.

So I built Fancy Cafe AI, which films a new version of the scene from your photo instead of pasting you onto someone else's clip. It doesn't host or copy the original TikTok video; every output is a newly generated scene.

How a 15-second scene comes from 1–2 photos

The flow is deliberately short:

1. Upload one or two photos. One person or one pet per photo, face clearly visible. JPG, PNG or WebP up to 10 MB. On a computer you can scan a QR code and send photos from your phone instead.
2. A free photo check runs right away, before any payment. It also spots pets automatically.
3. Pick roles. There are exactly two: On the table (sits on the tiny wall table, then realises and slides down) and Walks in (comes in and uses the table the right way).
4. Generate. You get one 15-second, vertical 9:16, 720p HD MP4 with cafe ambience sound, plus a GIF export and a share link. No watermark.

The casting rules were the most interesting design problem:

• One photo: you choose your role, and an AI-generated stranger, a fictional person, plays the other guest. You don't need a friend to make it.
• Two photos: photo 1 and photo 2 take one role each, and you can swap them in one tap before generating.
• Pets: cats and dogs only work in the On the table role, which honestly is where they belong. A cat sitting on the table and getting shown how cafes work is the best version of this meme.
• One person can't play both roles.

Under the hood, the videos are made with Seedance 2.0 (ByteDance) and Nano Banana 2 (Google). The point of using a video model rather than a face swap is that the AI generates your body, posture and reaction for this specific scene. It keeps your clothes, hair and glasses.

Full-body or half-body photos work best.

The audio choice was intentional too. The MP4 has cafe ambience only: espresso machine, cups, chatter, no music. When you post to TikTok, Reels or Shorts, you pick the trend sound from the app's library yourself. I've written a longer walkthrough with photo and caption tips in how to make a fancy cafe AI video.

No sign-up, on purpose

You don't create an account. You pay as a guest and the video arrives by email.

The honest limits

I'd rather you hear these from me than discover them at checkout.

It's paid, and there's no free trial. Every video uses real GPU time, so generation isn't free. The photo check is free, but there are no free credits. Pricing is one-time with no subscription: Single (1 video) $9.99, Trio (3 videos) $17.99, Party (5 videos) $26.99. Every video costs the same whether it has one person, two people or a pet, and credits never expire.

What happens when it goes wrong. If a generation fails, the credit comes back automatically. Unused credits are refundable within 7 days of purchase; used credits are not. If a finished video has a serious defect, such as a face that is clearly not you, a visibly distorted body, or someone missing, you can email support within 7 days, and after a manual review an approved claim gets one free re-generation.

Consent rules are strict. You can only upload yourself or people who said yes. No minors and no public figures, and misuse means a permanent ban. Your photos are sent to third-party AI services only to make your video, and you can delete the finished video and its photos from My videos at any time.

Results depend on your photos. Clear face, one subject per photo, no group shots. A blurry, crowded photo will give you a worse video, and no model fixes that entirely.

What's next

The plan is simple: add more viral meme templates beyond the fancy cafe scene, support more languages, and keep pricing one-time with no subscription.

If you try it, I'd love to hear which role you picked. I'm betting on the cat.