Katana AI Guide
The reference does the directing — these notes make the rebuild cleaner.
Start with the reference clip
Katana AI is reference-first: the uploaded clip anchors the motion, so the footage you pick matters more than any words. Four to fifteen seconds with one clear move — a walk, an orbit, a reveal — gives the engine a blueprint it can hold.
- One strong move per clip — stacked choreography confuses the rebuild.
- Stable framing beats handheld shake; the camera path carries over.
- Clear subjects in the source make it obvious what to replace.
- Use the preset library if you do not have a clip in mind.
Cast your subject photos
Subject photos tell the engine who or what belongs in the frame. One to eight images: more angles of the same subject help it hold identity through motion; unrelated subjects dilute it.
- Whole subject in frame, clear light — the same rules as a good portrait.
- Multiple angles of one subject beat single shots of many.
- JPG, PNG and WebP all work, up to 10 MB each.
- You need rights to every photo — and consent from anyone shown.
Optional audio and instructions
Attach one audio file if the edit should keep a specific sound — a track, a voiceover, a beat the cuts land on. The instruction field is optional steering: "keep the orbit, emphasize the product". Leave it blank and the reference decides everything.
- Audio is optional — silent MP4 exports are normal.
- Short instructions steer; long prompts fight the reference.
- Match the output aspect to where you will post — 9:16 for Reels and Shorts.
Reviewing the rebuild
AI video varies between runs. Watch the subject identity first — faces, logos, outfit details — then the motion against the reference. Keep the take where both hold, and label AI content where it matters.
- Check faces and product details frame by frame — drift shows there first.
- Compare timing to the reference: the cuts should land the same way.
- Re-run the same setup when the take is close; change one thing when it is not.