Tiny Desk AI video maker says a fictional stand-in stopped real faces leaking through

A maker of AI-generated Tiny Desk parody videos says it changed its workflow after tests found that editing footage of a real performer could leave traces of that performer’s face in the result. Its new approach edits a fictional AI-generated stand-in instead, a fix with useful implications for anyone making synthetic video from footage of real people.

Watch Desk analysis

What happened

In a 27 October Hugging Face community post, TinyDeskAI’s maker describes testing face similarity across generated frames. The creator says its earlier method sometimes retained features of the real coach shown in the reference clip, particularly when the uploaded person looked similar. The post discloses that its author sells a paid tool for the trend, so these are the maker’s own tests, not an independent assessment.

The reported change is to replace the real performer in the reference close-up with a fictional AI-generated stand-in, then edit that shot for each order. In the maker’s tests, all 151 sampled frames of one resulting close-up matched the uploaded test image and none matched the coach, according to the post’s face-scoring method. The creator says it also checks for failures that can slip through even when a generation returns a success status.

Why it matters

Synthetic video can make a person appear to perform or say something they never did. The creator’s account offers a concrete example of a less obvious risk: an edit intended to replace someone’s face may retain some of the original identity instead. A fictional reference may reduce that particular problem, though the creator’s test results are not a guarantee for other models or workflows.

The post also describes using a conventional face-swap model for some smaller faces, while noting that its scoring method can flatter face swaps because it measures identity features those systems are built to reproduce. In other words, a tidy number is not a substitute for watching the clip.

Our read

The practical lesson is to check who a generated face resembles, not merely whether the render completed. Using a fictional stand-in is a thoughtful safeguard in this maker’s workflow, but the evidence comes from the tool’s own creator and a specific set of tests. For viewers and creators, that is a useful warning with a sensible amount of uncertainty attached.

What to watch

• Whether the maker publishes further tests across different models and reference clips.
• Whether other creators find similar face leakage when editing real people.
• Whether synthetic-video tools make identity checks visible before a clip is shared.

Discussion spark: Should AI video tools require creators to check for traces of the original person’s face before sharing an edited clip, or should that responsibility remain entirely with the creator?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.