Tiny Desk AI

Tiny Desk AI is a paid video tool that drops 1 to 5 portraits into a fixed desk-concert clip, for people who want a 15-second video with generated sound.

Five roles in one fixed performance

The form has five parts: front backing singer, left backing singer, right backing singer, lead singer and saxophone player. You add a photo only for the people you want to replace. An empty slot keeps the original performer. Left and right follow your view of the video. The lead singer is optional. There is no prompt and no group photo.

The reference on the form is 13.6 seconds. Its labels put the three backing singers in the opening, 0 to 5.3 seconds, the lead in the middle, 5.3 to 11.6 seconds, and the saxophone at the end, 11.6 to 13.6 seconds. The finished file is 15 seconds, 720 by 1280, with sound. The FAQ says each person appears only during their part, and that framing and face size differ by role. Likeness, proportions, movement and lip sync can vary.

The image on this page is the site's own example, not a video made from here. It replaces all five roles with portraits the page names Gigachad, Harold, Roll Safe, Overly Attached Girlfriend and Salt Bae, and it says those portraits were created for the example. The clip beside them is labeled 15 seconds, 720 by 1280, with generated music and singing. The page says this one combination does not show every mix of roles.

The footer says Tiny Desk AI is independent, and not affiliated with NPR, NPR Music, Tiny Desk Concerts, any featured artist, or similarly named websites. A FAQ asks whether the video includes the original Outstanding recording. The answer says the output is generated music, singing and saxophone guided by the reference, not a promise of that recording or of voice cloning.

• A usable photo is one clear adult, face visible, as JPG, PNG or WebP up to 10 MB. The page says to avoid crowded frames, heavy filters, sunglasses and motion blur.
• Real people must be consenting adults, and you confirm you have the rights to the images. When you share the result, the footer says to follow the platform's rules for labeling realistic AI-generated video.

The listed price is $4.99, and checkout is described as not open yet

The pricing block says $4.99 in US dollars, one payment per video, tax included, with no subscription and no credit pack. That pays for one MP4 in My videos. A line under the button says about 17 minutes, and a 60-minute delivery limit. The estimate is labeled as one test on October 27, 2026, measured after upload. The 60 minutes start at confirmed payment and include upload time. After that limit, the page says you can ask for a full refund.

The how-it-works step still says to sign in and pay once when ordering opens. The privacy page says the production provider will be named before checkout opens. Sign-in is Google, or an email and password. This page was not taken through payment.

A failed or unusable video

The refund page splits the cases.

• If making the video fails, they redo it once automatically, or refund in full immediately if you prefer. A failed redo is also a full refund.
• Unusable means the subject is unrecognizable, falls back to the reference performer, misses the promised performance, is severely distorted, does not match the sound you paid for, or will not play. Tell them within 7 days of delivery. They say they decide within 2 business days, then finish a redo or start a full refund within 1 business day. You pick which.
• If the photo cannot be used, you can swap another photo once for free, or take a full refund.
• After the delivery time shown before payment, you can cancel for a full refund. A late file does not remove that option. A full refund returns the amount actually paid, including tax, with no processing fee taken off. Starting the refund and the money arriving are separate.

The photo stays local until payment

Until payment succeeds, the photo stays on the device. The browser keeps a local draft for up to 24 hours, and the form can remove it. After payment, a standardized copy goes to private storage, and the generation provider gets the photo plus template instructions. An automatic description of the subject is stored with the photo so a retry can reuse it. Payments go through Stripe, with Waffo as a backup. The site says it does not receive the full card number.

• Their photo copies and those descriptions are deleted within 7 days of delivery, cancellation or refund. The video stays downloadable for 30 days. Delete removes access immediately and queues deletion of their copies. Other customers cannot open your files, and there is no public gallery. Buying a video does not give permission to use it as a public example.
• They name Atlas Cloud as a processor. Its published policy limits its own use of customer content for training unless you authorize it in writing, and its standard retention can run to 7 days. The privacy page says third-party model providers have separate terms, that this site has not verified a no-training promise for the upstream routes used for recognition and video, and that it does not promise one. The service is not set up for Enterprise Zero Data Retention. Deleting their copy does not delete a copy held by a provider, or a file already downloaded.
• The operator line says Tiny Desk AI, and the public contact is [email protected]. It does not name a person. Analytics is off unless you allow it. Google Analytics then gets browser and device data, a cookie visitor id and sanitized paths, not the name, email, photos or card. That choice lasts 45 days.