HeyGen breaks down its 16s one-take AI video: pin first and last frames, model fills the middle
HeyGen · x · 2026-09-12
HeyGen shares the full recipe behind its 16-second one-take AI video where a cardboard box bursts into a fully furnished living room — no 3D, no compositing, no post.
Inputs
- Two stills: an empty room start frame, and the same image edited to add furniture as the end frame (not a fresh render)
- A custom avatar so the person is the author, not a model-invented stranger
- One prompt, using Seedance at 16:9, asking for 15 seconds
Core method
- One image + prompt lets the model improvise the ending; pinning both frames locks the payoff — the model's only job is inventing the middle
- Risk shifts: the payoff shot is a still you already signed off on; all invention happens in the seconds where furniture is airborne
Output: 1920x1080, 23.976fps, H.264 15 Mbps, 16.27s.
More from Multimodal
- Nex-N2.5 Pro, a 397B multimodal model focused on Computer Use, quietly lands on OpenRouter — nikola_mr64990 · 2026-09-12
- Can't Disable Audio Generation in H3? User Seeks Workaround — Obvious-Leg-5604 · 2026-09-12
- Score-Conditioned Stable Audio Now Running, New Way to Steer Music Generation — matdryhurst · 2026-09-12
- GPT Image 2.5 and Seedance 2.5 land in CapCut for AI ad workflows — HeyNayeem · 2026-09-12
- ComicForge turns written stories into finished comic books in 29 languages — testingcatalog · 2026-09-12
- One prompt makes EDM rave visuals of sound vibrating sand and bioluminescent plankton — jfischoff · 2026-09-12