GPT image 2.5 nails character consistency in multi-agent video restyle workflow

kagigz · x · 2026-09-10

A demo shows GPT image 2.5's impressive character consistency, with the full pipeline disclosed: agents (Astra, Sunburst, Codex) sampled a 30s video into 249 frames, picked an ink style plus 7 reference images, prompted Sunburst to keep pose from the original while borrowing style from references, had sub-agents review outputs, and used FFmpeg for timing, audio, and side-by-side rendering.

Original post →

More from coding & agent

coding & agent channel →