AI by department · Design, look and the camera department
Concept is where practice is furthest ahead of policy. The camera department owns almost none of it.
Image models are in daily use across concept, look development and costume ideation. Almost nothing touches the camera itself: nothing in this category has timecode, genlock or lens metadata. What has changed is how a shot gets agreed before the day, and what post is permitted to do to your images afterwards.
Last checked September 2026. This page absorbs what used to sit on a separate camera page. Nothing on it is legal advice.
Three phases
Three phases, and the camera department only owns one of them.
| Phase | What is in use | Who controls it |
|---|---|---|
| Before the shoot | Previs, look development, generated reference, lookbook assembly, breakdown into a shot and equipment list. A director and a cinematographer arguing about a shot the week before rather than on the day. | Usually production or the director. The department inherits it as though it were a decision. |
| On the day | Almost nothing. Some monitoring and framing assists, and automatic logging of takes and metadata. | You. |
| After the shoot | Reframing, denoise, upscaling, object removal, relighting and face work. DaVinci Resolve 21 ships a Face Age Transformer, a Face Reshaper, CineFocus and UltraSharpen. | Post. Where a look agreed on the day quietly becomes a different look. |
Reshaping a face is a change to a person’s appearance, so it is a consent question for the performer as well as a scope question for the cinematographer. Ask at prep what post is permitted to do to the images, get it into the prep notes, and name who signs off a change to the look.
Operator filming a fire effect at an outdoor event
What it does badly
What it does badly.
- A frame is not a drawing. No elevations, no dimensions, nothing construction can cost or build.
- It has no idea what a set is for. Doors that cannot open, corridors with nowhere to put a light.
- It cannot hold the same room twice, and the art department’s product is one location from twelve angles.
- Period and cultural accuracy is pastiche, and costume gets almost nothing: no pattern, no fabric weight, no fitting, no multiples.
Text and signage come back wrong more often than people check, and generated graphics can reproduce real trademarks and real people.
The exposure, and the line
On the cheap plan you are publishing the designs, and here is where we stop.
Midjourney’s default is public. An art department generating unreleased designs on a basic or standard subscription is publishing them; privacy, Stealth Mode, starts at the Pro tier, $60 a month on Midjourney’s published prices in September 2026. Under a non-disclosure agreement the cheap plan is arguably already the breach.
We use these tools, teach them and test them. Do not prompt with a living artist’s name or in an identifiable individual’s style: training a model in another artist’s style without cleared rights is the first thing Netflix names as needing written approval. Do not pass a generated board to construction or costume as though it were measured.
- That a generated concept pack saves an art department money once drawing, costing and construction are counted
- That any current image model is safe as a final delivered asset without human work on top
- That the output is automatically yours, or that a vendor’s commercial-safety claim covers the third-party models it resells
- That a generated previs frame is a promise to the director that the shot is achievable
The skills that make you harder to replace: draw over rather than hand over, because UK copyright requires a human author; measurable 3D; draughting and set standards; and a provenance log per asset recording model, version, settings and licence tier.
