PreCoG: Preview-Edit for Efficient Video Correction in Generation-Time
Published in Manuscript, 2026
PreCoG reduces the cost of iterative video correction by assessing low-cost previews, reusing cached generation trajectories, and fully rendering only the selected candidate. PreQA diagnoses unmet visual requirements and constructs corrective conditions. PreEdit is a training-free editor that accounts for source-condition misalignment: a generated trajectory may only partially satisfy its condition, so correction must address unmet requirements while preserving satisfied ones. It uses a target-counterfactual prediction contrast and blockwise soft projection to guide source-relative updates and limit unrelated changes.
