A correction technique for AI video generation: when a generated video comes out flawed, resend the video itself — not just a rewritten prompt — to a model with built-in video-understanding capability, along with specific written critique (concrete notes on shot order, pacing/tempo, missing sound effects, animation timing), and ask for a regeneration.
This turns a failed generation into a conversational fix rather than a manual re-edit: the model watches the actual flawed output and corrects against the real mistake, instead of the operator guessing blind at what went wrong in the original prompt.
Demonstrated with Higgsfield Supercomputer's built-in Gemini video understanding, used to fix a botched multi-shot commercial video produced by a custom skill (Higgsfield's 'C-Dance 2' multi-shot prompter) after a rushed, carelessly-read prompt produced flawed output — sending the bad video back with critique produced a corrected regeneration.
Related: Take Compositing for Best Performance (compositing across multiple takes rather than regenerating from critique), Iterative, Non-Linear Shot Generation ("Nothing Is Locked Until You Lock It").
Из тем: Holding Continuity Across Independently Generated Shots