Tiny Desk AI

Building a one-photo generator for a 7-second meme: the choices behind Bicep Hearts

Zaire Prow — Wed, 04 Nov 2026 16:53:43 GMT

Some trends are hard to describe. The bicep heart is not: flex one arm, curl the other hand against the bicep, and the two shapes make a heart. It took off on TikTok in late October 2026, and within days I wanted a version that didn't require filming, good lighting or arms anyone would call impressive. That became Bicep Hearts, which turns one photo into the clip.

This post is less about the trend and more about the product decisions: what I kept, what I refused to build, and why.

The pipeline, in two steps

A single photo isn't enough to animate directly with a consistent result, so generation is split:

Reference image. Nano Banana (Google's Gemini image model) turns the upload into a clean reference image of the subject in the pose context.
Video. Seedance 2.0 (ByteDance's video model) animates the bicep heart from that reference.

The output is fixed on purpose: 7 seconds, vertical 9:16, 720p HD, MP4 plus a GIF, no watermark. Fixed specs mean I can tune one thing well instead of many things badly.

Arm levels instead of a prompt box

Users don't want to write prompts for a meme. So instead of free text there are three arm levels: Normal (your own arms, a little more defined), Buff, and Extreme. Extreme exists because a large share of the joke is how absurd you're willing to go. Three choices are enough to cover "me, but tidier" through "obviously ridiculous."

The one-subject rule

Every video has exactly one subject. No two people, no person plus pet, no two pets. This was a deliberate cut. Multi-subject scenes multiply the ways a pose can go wrong (whose hand is whose, which arm flexes), and the trend itself is a solo format. Saying no up front is better than charging for a clip with tangled limbs.

Pets got their own beats

Pets were the most fun part to design. A dog can't really make a heart, and that's the point. The pet version has its own choreography: stand on hind legs, flex one front leg, curl the other paw into a rough heart, hold the pose with a dead-serious face, in a bright, clean living room. The animal keeps its breed, fur colour, markings and face. Cats and dogs work best. Pets don't get arm levels; there's no "Extreme" golden retriever. More detail is on the pet page.

Checking the photo before taking money

The worst experience is paying for a video that was never going to work. So a free photo check runs right after upload and before payment. For people, the photo must be waist up with both arms visible; files are JPG, PNG or WebP up to 10MB. On desktop you can scan a QR code and send the photo from your phone, since that's where the photos usually are.

Privacy defaults

For a tool that handles faces, I wanted the boring answers to be the default ones:

Photos are used only to generate the video.
Photos and videos are deleted automatically within 7 days, or immediately from My videos.
Nothing is used to train AI models, and no face templates are built.
Only upload yourself, an adult who agreed, or your own pet. No minors, celebrities, public figures or strangers. Abuse means a permanent ban.

No account, no subscription

You don't sign up: upload, pay, and the video is emailed to your checkout address (Google sign-in is optional if you want an account). Pricing is one-time: $9.99 for one, $14.99 for three, $29.99 for ten. Credits never expire. There's no free trial, which I'd rather state plainly than hide. If generation fails, the credit returns automatically; unused credits are refundable within 7 days; and if a finished video has a serious defect (wrong face or pet, obvious distortion), support reviews it and grants one free re-generation. Full breakdown on the pricing page.

Silent by design

The video has no audio. Most bicep heart clips use an original sound from @trendformusic, and that sound isn't mine to bake in. You add it in TikTok or Reels when you post.

What I'd tell another maker

Pick a format narrow enough to describe in one sentence, fix the output specs, check inputs before payment, and make deletion the default. Most of the work was deciding what not to support.

Bicep Hearts is independent and not affiliated with TikTok or any creator of the trend.

Twelve seconds of somebody dancing in a lit window

Zaire Prow — Sat, 31 Oct 2026 12:19:07 GMT

The best short video format of the last week is also the simplest one to describe: a phone, at night, from across the street. Dark building. One window still lit. A zoom that is a little too eager. Somebody dancing alone next to a small speaker, who then notices they are being watched, grins about it, and closes the curtain.

That is the Window Dance trend. It landed on TikTok on 22 October 2026 as a CapCut template and it is the dance version of the Window Peeper meme from two days before, where the person in the window is eating a burger. From 25 October the pets arrived, and the pets are better, because a dog does not grin when it notices the camera. It stops moving and stares.

The template has one limit that started to bother me: it swaps your face onto a single fixed clip. Same room, same clothes, same moves, everybody. So I built Window Dance AI to generate the entire shot from one photo instead.

What the model actually has to get right

The hard part is not the dance. It is the camera. The shot only works if it reads as a phone held by a person hiding across the street: a bit of drift, a zoom that overshoots, the window as the only light source, the rest of the frame falling off into black.

So the pipeline keeps the subject fixed and lets everything else be invented. A free AI photo check runs first, before payment, and rejects minors and obvious celebrities, because the premise depends on the person in the window being a consenting adult. Then the photo becomes illustrated reference images that hold the face and the clothes while Seedance 2.0 builds the building, the street, the room and the light around them.

Output: 12 seconds, 9:16, 720p, MP4 plus a GIF, no watermark. The audio is an original instrumental with room sound, deliberately not the trend's TikTok original sound, because that audio belongs to somebody else. You add the trending sound in TikTok when you post.

The boring parts, in public

One-time credit packs, no subscription: 1 credit $9.99, 3 credits $19.99, one credit per video, credits never expire. A failed generation refunds its credit on its own. Unused credits can be refunded within 7 days. There is no free tier, and I would rather say that up front than hand out a trial that turns into a paywall.

One person or one pet per clip. No duo, no person plus pet. Everything is AI, and the label should stay on when you post it.

When you should use the CapCut template instead

Honestly? If you want your face on the standard clip and you already have CapCut installed, the template is free and takes ninety seconds. Use it. The generator earns its price only when you want the scene rebuilt around your own photo, your own clothes or your pet. I wrote that comparison out properly instead of pretending there is no alternative: CapCut template vs AI generator.

Pets: windowdance.com/pet-window-dance The full trend breakdown, with timings: windowdance.com/how-to-do-the-window-dance-trend

Not affiliated with TikTok, CapCut or the creators of the original videos.

I built a Tiny Desk AI generator that turns one photo into a 15-second concert video

Zaire Prow — Sat, 31 Oct 2026 09:24:18 GMT

A friend sent me the Charlie Wilson 2023 Tiny Desk Concert clip and I kept thinking: everyone should get to be in that band.

So I built https://tinydeskai.studio . You upload one to five photos. The first person sings lead, the second plays sax, the rest sing backup, and any empty spots get filled with AI musicians. About a minute later you get a 15-second vertical 720p MP4 with the stage audio, plus a no-music cut for platforms that flag copyrighted audio.

Some things I learned building it:

Multi-person identity consistency is the hard part. Keeping the same face across a wide, moving stage shot meant generating a reference portrait per person first, then conditioning the video generation on those, rather than trying to do face swap afterwards.
Audio sync matters more than resolution. People forgive 720p, they do not forgive the singer's mouth being off by 200ms.
A no-music export is not optional. TikTok and Reels muting the whole clip is worse than shipping two files.
Guests should not have to make an account. It is a guest checkout at $7.99 per video; adding a signup wall cut conversions badly in testing.

If you try it, tell me what breaks. Especially curious about group photos with more than three people.