5-second chunks + last-frame chaining beats one-shot 15s AI video generation
Downtown_Way5997 · reddit · 2026-11-12
After a week of failed attempts at 15-second clips, this author switched to a chunked workflow that works far better:
- Generate every shot as a 5s clip, grab the last frame, touch it up if needed (hands, artifacts), and use it as the start image for the next 5s chunk — same prompt style, only the action changes.
- Why: past 5s their GPU starts swapping to system RAM and one job goes from 15 minutes to over an hour; three 5s jobs finish much faster, and a bad chunk means redoing only that piece.
- Face drift drops because every chunk starts from a frame already inspected.
The catch: seams — motion can jump at cuts, so they keep camera moves simple (slow push or static) and cut on an action; color drifts over 3-4 chunks, but a quick color match in the editor fixes most of it.
More from Multimodal
- Creator teases a Honor of Kings remake made with Fable 5.2, due tomorrow — FuSheng_0306 · 2026-11-12
- Pushing Midjourney to obey prompts: low --stylize plus --raw, with film-portrait examples — michaelrabone · 2026-11-12
- AI extends human performances, not replaces them: Jon Erwin on filmmaking with Seedance 2.5 — LudovicCreator · 2026-11-12
- Stranger Things reimagined in 1972 style with AI video — techhalla · 2026-11-12
- Action Director ships as standalone 3D posing app that live-links back into ComfyUI — shamomylle · 2026-11-12
- Ditch "cinematic" prompts: 4 physical camera rigs that break the AI gloss, with exact prompt lines — Signal-Chipmunk-9634 · 2026-11-12