Make app demo videos with Veo, then switch to motion transfer from your own footage when it fails

A small app that teaches physical movements needs a clear demo video for every move. Hiring an athlete and a film crew is expensive for a solo founder, and generic stock footage does not cover niche moves.

1. Before building, write hypotheses and design analytics events, then write a concept document and screen specs (an AI design and coding tool rebuilt the Flutter interface across 18 screens). 2. Create a starting still of a reference athlete with an image model (gemini-3-pro-image). 3. Try text-and-image video generation with Veo 3.1 through the Gemini API, describing each movement in a JSON spec and generating with a script. This worked for standard standing moves. 4. For moves Veo could not produce (hip escape, reverse shrimp, sit-through, 90/90 stretches), film yourself doing the move and use Wan 2.2 Animate on fal.ai to transfer the motion onto the reference athlete. 5. Clean up the output with OpenCV, MediaPipe and ffmpeg: stabilise, grey out the background, interpolate frames for slow and fast versions, repair freeze frames, and crossfade loop seams. 6. Generate all spoken cues in one batch with Gemini text-to-speech, compress them to small audio files, and bundle them so they work offline.

Mixed. The app is live with AI-made demo videos and voice cues. Veo failed on floor-based moves even with detailed prompts, and raw motion transfer needed heavy post-processing (for example a freeze frame roughly every 0.8 seconds). Early users reported that some move videos look wrong, which the author lists as the top open item. No costs, user numbers, retention or revenue are reported.

No costs or business results are reported. The specific Veo preview models used have since been retired by Google, so the exact setup cannot be repeated. One solo builder, one niche app. The author offers help with AI pipelines by direct message (no fee stated) and links to the app. Steps are described but no code is shared.

Gemini API access, a fal.ai account, the ability to film yourself performing each movement, and comfort with Python, OpenCV and ffmpeg.

• DEV Community (toshipon, BUILD GEEKS) ↗ Firsthand write-up, published October 26, 2026

Source published: October 26, 2026. Last reviewed here: October 28, 2026. Spot a mistake? Tell us.