Building a One-Photo Concert Tool: The Work Around the Video Model

王瑞 — Mon, 26 Oct 2026 12:49:05 +0000

I built AI Tiny Desk, a paid one-photo concert generator for the Tiny Desk AI format. The site runs TypeScript with Hono on Cloudflare Workers, D1 for task and credit state, private Backblaze B2 storage, and PayPal/Waffo checkout. This post describes the implementation we actually shipped, rather than substituting a more familiar Next.js/R2/Stripe stack.

One input, five roles

The user uploads one consenting adult portrait. That person becomes the lead in two shots; the prepared band retains four fictional identities. Separating dynamic shots from fixed band shots reduced the number of already-approved sections we needed to regenerate.

Time in accepted sample	Shot	Identity rule
0–1.375 s	Opening singer	Uploaded adult portrait
1.375–6.667 s	Keyboard and backing vocals	Fixed fictional band
6.667–13 s	Singer close-up	Same portrait as the opening
13–16.375 s	Saxophone; keyboard in corner	Keep these two people distinct
16.375–19.042 s	Backing group	Reuse the prepared band

Reliability spans several independent states

A provider submission is not a finished file. A finished file is not proof of a correct face. A restored credit is not a cash refund. Our UI and support process keep these distinctions visible. Users sign in with Google or an email code, buy one-time credits, and reopen their result from their account.

The concert route uses Seedance 2.0 Mini through EvoLink. The model output is silent; browser-side Mediabunny remuxing prepares the sound copy. We keep the silent download available, including when sound preparation fails.

A small test log, not a benchmark

Check on October 26, 2026	Observed result	Evidence scope
Mixed lead reference, Seedance 2.0 Mini	Close-up replaced; opening still showed the old lead	Developer visual review; not a pass-rate study
Isolated opening, first submission	Duration validation failed	Provider status recorded failure; no output
Isolated opening, corrected reference	Completed in 116 seconds	Provider reported duration for this one task only
19-second Kling O3 edit experiment	Rejected for input duration	Failed experiment; not our production model
Browser audio remux	459 encoded video packets unchanged in the production fixture	Local accepted-sample comparison; not a buyer transaction

Audio alignment needs a known offset

The complete operator-provided source is aligned from the opening shot, while the video packets remain unchanged. This command illustrates the same sequencing for a local file you are permitted to use:

# Use your own permitted source audio, aligned to the video start.
ffmpeg -i supplied-video.mp4 -vn \
  -af "apad=whole_dur=19.25" \
  -t 19.25 -c:a aac -b:a 160k aligned-audio.m4a
ffmpeg -i silent-render.mp4 -i aligned-audio.m4a \
  -map 0:v:0 -map 1:a:0 -c:v copy -c:a copy \
  -shortest -movflags +faststart output.mp4

One practical lesson: measure the input that the provider actually receives. Inspect duration and stream metadata before spending another paid attempt. Another is to inspect the opening and close-up independently; a convincing face in the middle of a clip can hide a failed first shot.

The prompt guide and creation tutorial document the workflow without exposing private inputs, task URLs, or customer information.

Product facts and limits

AI Tiny Desk uses ByteDance Seedance 2.0 Mini through EvoLink for the concert workflow. Standard generation is 480p. The output is approximately 19 seconds in vertical MP4. One credit makes a Standard concert; one-time packs are $9.99 / 1 credit, $19.99 / 3 credits, and $29.99 / 5 credits. Credits do not expire. Confirmed generation failures restore credits, while cash refunds and serious visual defects have separate refund terms.

The current sound version uses performance audio supplied by the site operator including the complete original opening. A silent copy remains available. Buying credits does not grant music rights. We support one consenting adult per input; we do not advertise pet or child support. We do not guarantee exact likeness, a render deadline, search rankings, or that an uploaded photo will pass a model's safety checks.

Sources and disclosure

We built and operate AI Tiny Desk; this is our own development write-up, not an independent review. Test portraits and our demo band are fictional adults. The tool is independent of NPR, Charlie Wilson, The Gap Band, TikTok, and CapCut. AI assisted the preparation of this write-up; technical statements were checked against our source code and saved test artifacts.

The original performance is Charlie Wilson's NPR Music Tiny Desk Concert, published June 5, 2023. The artist's official website identifies Outstanding in the set. This establishes the performance reference, not the identity of the first AI remix creator.