Systematic stress testing of intermediate-keyframe error propagation

Conduct a systematic stress test of how hallucinated details, incomplete edits, and other errors in the intermediate image-editing stage propagate through Chain-of-Editing to all output video frames, particularly for geometric transformations and novel-content synthesis.

Background

VINCIE-NExT uses the edited keyframe produced by the intermediate image-to-image stage as a hard spatial anchor for subsequent video generation. Consequently, errors in that keyframe can be propagated rather than corrected across the generated sequence. The paper reports that self-editing is weaker for geometric transformations and novel-content synthesis, while stronger external image editors improve results, but it does not systematically characterize the failure risk across these conditions.