Contents
In brief
A two-person team could not hand-cut short product videos on schedule. A Dev.to write-up covers wiring Runway Aleph: the model makes a bounded edit to real footage instead of inventing a new clip. Text-to-video demos well and fails continuity in production.
What happened
Generate-first tools produce independent clips: “the same” product is a different product each time. An in-context model takes source video and a scoped change — remove a price tag, keep box, surface, lighting.
Pipeline: raw footage → JSON shot manifest → edit passes → human QC → publish. The manifest was underestimated: ops are replayable, one per entry. Batching “remove / relight / add a plant” ruins all three and cannot be debugged.
Explicit preserve clauses stop quiet background re-renders. Similarity auto-QC catches gross failures and misses semantics: fingers, gibberish signage, broken reflections. A content-addressed cache on source + prompt cut more than half of iteration spend.
Why it matters
The win is not “impossible became possible.” Expensive work got cheap enough to attempt. Real cost sits in full-resolution review, not generation seconds.
In practice
- Add one stage first (object removal), don’t rebuild the whole pipeline.
- One operation per pass; batch only if you accept not knowing which clause broke the frame.
- Always state what to preserve, not only what to change.
- Human QC on risk regions at full resolution, not in a preview pane.
- Cache identical source+prompt pairs; benchmark including review time.
Takeaway
AI video in content works as controlled editing, not a factory of random clips. Manifests, narrow passes, and human QC matter more than the model brand.