Lore
Turning image variations into talking UGC video ads
The video argues that static ads are good but video ads perform better. The second half of the pipeline therefore turns each generated character image into a realistic talking video.
How the variations are used
- Each variation can become its own video ad.
- Several variations can be stitched into one compilation.
- The character stays consistent across every video, even as outfits, backgrounds and products change (Consistency as the trust signal in ad campaigns).
Inputs
After generation
The visuals are only half of what makes the video pass as real. See audio-realism-as-last-step-for-ai-ugc.
Repeat-and-stitch across variations
Every variation image from the Spiel image-variation step (Spiel Magic Edit: variations from one source image) is reused as the input to the UGC video step, so one character yields a whole set of videos.
- Repeat the UGC builder process once per variation. Each pass keeps the same character and product but gets a new outfit and background.
- Turn each result into its own video ad, or stitch the clips together into one longer video.
- Direct each clip with an action prompt, see avatar-action-prompt-direction.
- Finish each clip with the audio treatment in elevenlabs-voice-changer-on-extracted-audio and elevenlabs-subtle-background-ambience.
The presenter says they make dozens of these videos every day. The claim is that a campaign can be built "fast, cheap, and at scale without ever having to book a shoot". The source also states that "video ads perform way better" than static ads.