references/editing

Place cuts based on motivation: cut on action, on a look, on sound, on an idea, on emptiness, or on impact. Every cut must advance information, shift attention, hand off motion, reveal a reaction, or serve emotion; unmotivated cuts read as decorative and should be removed.

Fit a beat grid by running librosa beat tracking on the BGM and least-squares fitting the beat sequence to get beat length and phase. Pin all cuts to integer beats, then re-measure the rendered audio and verify every cut lands within 3 frames of its designed beat.

Feed and vertical content typically uses a 0.7–2.0 second average shot length, with hooks allowing 0.2–0.5 second punctuation cuts. Information segments should return to 3–6 or 5–12 seconds so ideas have time to land, and the median shot length should sit below the mean.

Use a dissolve only as an honest repair for continuity gaps that cannot be fixed, such as mismatched light levels or shifted textures. Hard cuts are the default; dissolves consume running time and imply time passage, so never dissolve between two shots continuous in time.

The slideshow feeling comes from the cut, not the footage. Enforce the eight anti-slideshow rules: overlap motion across cuts, break shot size or angle by at least two rungs, keep one continuous audio bed, vary shot lengths, and never open two consecutive shots on static frames.

Generate 1.5 seconds longer than the needed cut for simple shots and 2.5 seconds for character performance or dialogue. Discard the first 8 frames and last 12 frames of every generated clip before evaluating content, since heads ramp in and tails drift.