← sources

A neat trick for generating longer AI videos: working backward from the end

source: reddit (reddit:StableDiffusion)published 2026-10-29 21:04ingested 2026-10-30 02:38

retained text

I’m testing a way to reduce quality drift in long AI videos. Normally, each new segment continues from the previous one, so errors can accumulate like a game of broken telephone.

Instead, I generate the opening first, then the ending using the opening as a reference. From there, I work backward toward the middle, keeping the original image and voice as references. The last segment generated bridges the opening to the backward-built sequence.

For a six-part video, the generation order is **1 → 6 → 5 → 4 → 3 → 2**, but the final video plays **1 → 2 → 3 → 4 → 5 → 6**.

document fetched 2026-10-30 00:32 · canonical https://reddit.com/r/StableDiffusion/comments/1xiw6jo/a_neat_trick_for_generating_longer_ai_videos