Give up on long AI video: 5s chunks + last-frame chaining works far better
Candid-Pay-9023 · reddit · 2026-11-12
After failing to generate 15-second clips in one job on a local GPU, the author switched to a chunked workflow:
- Generate 5-second clips, grab the last frame, fix small artifacts (hands etc.), and use it as the start image for the next 5s chunk; keep the prompt style consistent and only change the action.
- VRAM reason: jobs beyond 5s spill into system RAM, going from 15 min to over an hour; three 5s jobs finish much faster than one 15s job.
- Benefits: redo only bad chunks, and faces drift less since every chunk starts from an already-checked frame.
- Pitfalls & fixes: motion can jump at seams — keep camera moves simple (slow push or static) and cut on action; color drifts over 3-4 chunks, fixable with a quick color match in the editor.
The author asks the community for cleaner ways to hide the seams.
More from Multimodal
- Moss Image Team Hits 6 Million Hugging Face Downloads, Teases Anime and Edit Models — JadedScorpion · 2026-11-12
- Stunning AI-generated fantasy short: knight vs lich scene looks film-grade — PrinceMilktongue · 2026-11-12
- One Midjourney --sref code, three consistent rainy-day illustration styles — sergeantsref · 2026-11-12
- Seedream 5.5 turns one cat idea into a full brand-ready IP universe — ZabihullahAtal · 2026-11-12
- Hands-on with MiniMax H3 Image: intricate 4K art in 10 seconds, 5 example prompts — HeyZoyaKhan · 2026-11-12
- KENDRICKS P.I. E01: A fully AI-generated film noir detective short — Terminal0854 · 2026-11-12