AI Film Shot Board

This is the rule that costs the most when broken. "Outward" is meaningless to a model that has not decided where inside and outside are. "Away from camera" cannot be misread.

Never write forward, back, out or in. The six coin flips to check before any prompt goes out:

1. Direction — toward camera, away from camera, or across frame left to right.
2. Hinge and swing — doors, laptops, lids, car doors. Name the hinge edge and which way it opens relative to camera. Say what it does not do.
3. Which hand — and does it match the shot before? Remember the mirror: facing camera, someone's left hand reads at frame right.
4. Camera side — which shoulder you are behind, which side of the vehicle. Get it wrong between two shots of one scene and the geography flips on the cut.
5. What must NOT move. Naming the still thing is as useful as naming the moving one.
6. What is not in frame — crop points, and anything that must stay out: a second person, a head entering from above, text, signage, a watch.

If a shot carries an emotional arc rather than one state, write it as timed beats in the ACTION block — 0.0–1.0s, 1.0–2.0s — and say which beat is the turn.

Image-to-video starts from the first frame. Anchor the still on the emotional destination and the shot has nowhere left to travel — you get a flat, held expression, which is the single most common failure in generated performance. Anchor on the first beat instead and let the motion prompt carry the rest.

The reference sheet is a document. These are what actually keep a character consistent:

1. One canonical hero portrait per character, uploaded on every generation. Never chain off the last image you made — drift compounds generation over generation. Always go back to the original file.
2. A fixed identity block, pasted verbatim every time. Not paraphrased.
3. Person and setting described separately. The identity block never changes; everything below it does.

The reference set: front (the casting — run it 4–5 times, pick one, discard the rest), three-quarter, profile, full body, back, seated, hands (generated from the full body so sleeves carry over), plus an expression still for any specific look and a speaking still if a mouth moves on camera.

Expression and speaking references leak. Load them only for the shots that need them, or that look turns up in every frame the character appears in — unless a shot deliberately travels between two expressions, which is the one case for loading both.

Props are not wardrobe. A hat someone carries but never wears goes in the scene prompt. Put it in a reference and it appears in every shot. New wardrobe means a new named variant with its own portrait — ad-hoc edits are how one character quietly becomes two.

Stills are cheap or free. Video is not. Open the anchor still and check it against its panel before it goes to the video tool:

• Right person, matched against the reference side by side at full size
• Wardrobe correct — every item, right colours
• No accessories that should not be there
• Direction and handedness right, and consistent with the shot before
• Nothing cropped that should not be, no letterboxing baked into the file
• Nobody in the background who should not be there

Every credit lost from here should be lost to motion going wrong — never to something already sitting in a frame you approved.

When a still is wrong in one detail, upload it on its own and ask for the single change:

"Keep this image exactly as it is — same person, same wardrobe, same location, same lighting, same composition, same grain and grade. Change one thing only: [the fix, in camera-relative language]. Nothing else about the image changes."

You keep the frame you already approved. If it drifts the face or the light while fixing the detail, fall back to a clean generation.

Putting a duration in the prompt just gives the model something to argue with. Generate long, never short, and trim in the edit so you have handles to cut on.

Which also means: make sure the beat actually happens inside the part you will use. A gesture that lands at 6.8s of a 7s clip is gone the moment you trim to four.

Identify the one or two shots most likely to fail — usually a lit unobstructed face, sustained on-camera speech, a precise hand action, or any shot with an emotional arc. Take those from reference to still to finished clip before generating anything else.

Half an hour there tells you whether the whole approach holds. The alternative is finding out after you have built everything else.

Expect 8–12 generations per usable clip. If that number is frightening at your tool's prices, that is the signal to move iteration somewhere cheaper — not to plan for fewer tries.

You cannot judge a generated clip by watching it once. Extract frames and check identity against the reference at full size, wardrobe item by item (watches and rings appear out of nowhere), whether the beat happens inside your usable window, direction and handedness against the neighbouring shots, the background, and the framing against what the panel specified.

Motion quality lives between frames, so sampled stills can hide it. Warping, morphing, jitter and drifting fingers need watching at speed.