Previz Never Made Sense for Photography. Until Now.

Previsualization has always been a film and TV thing. And for good reason. When you have a hundred person crew, a six-month pre production window, and VFX that need to be designed before a single frame is shot, spending the time and money to build the shot before you build the shot makes sense.

Photography however, never really got there. Not properly. What passed for previz on a campaign was a mood board, a reference deck, maybe some pulled images from shoots that may or may not have anything to do with your product, your lighting or your brief. Directionally useful. Creatively suggestive. But not really previz. More like an educated guess dressed up in a nice PDF, which in many cases still works just fine. But when the stakes are higher, AI changes the math.

Genuine previsualization for stills production is now viable in a way it simply wasn’t before, not because anyone redesigned the concept, but because the speed and cost no longer make it impractical. You can now test backgrounds against actual product. Validate lighting treatments before the studio is lit. Explore environments without building anything. And do it fast enough that it fits inside a production timeline.

The other thing that changed? Previz no longer has to stop when the shoot starts.

What actually happens on set

This is starting to show up on large scale productions, not as a proof of concept, but as a working part of the day.

The setup looks something like this... a remote AI team working in parallel with the shoot, exploring directions in real time, with someone on set relaying between that team and the photographer or director. A question comes up. The remote team tests it. Whoever’s running point on set brings back something the director can actually look at and react to, in minutes, not after wrap.

This matters most when motion is involved, since the cost of getting it wrong in post is so much higher. But it holds up for stills too, particularly on larger set builds, complex backgrounds, or anything with a group of talent where there’s more that can go sideways. Most brands already run their own previz internally for sign off before a shoot. This is the same instinct, just moved onto the floor, for the moments when a question comes up about how to actually execute something.

The thing this solves is a very old and very expensive problem. Decisions that get deferred to post cost more in post. They cost more in time, in resource, and sometimes in the relationship with the client when something that should have been resolved on the day surfaces three weeks later as a problem nobody wants to own.

What AI allows is for those decisions to be made at the right moment, with the right context, while the option to actually change anything still exists.

This works because inference is fast. Feed the model a reference and it hands back a usable composite in seconds. The model doing the work matters more than the speed does.

Why a generalist model is not enough

Most AI tools that have made their way into the production conversation run on foundation models. Large, general-purpose architectures trained on enormous datasets. Impressive, genuinely. Also, by design, built for nobody in particular.

For brand production, that’s the problem. It’s never seen your product, doesn’t know your packaging, has no idea how your materials actually behave under real light. Ask it to generate or composite around your product and it’s pattern matching against data that has nothing to do with your brief. The result looks like a guess, because it is one.

This is where fine tuning comes in. LoRA, which stands for Low Rank Adaptation, is a method of fine tuning a large model by training a small set of additional weight matrices alongside the existing architecture. You are not rebuilding the model. You are adapting it so that it understands something specific. A product, a material, a visual language, a brand.

The practical difference is significant. A model trained on a client’s product understands that product at the level of its geometry, its surface characteristics, how it relates to light. The outputs are not approximations. They are decision-ready. On set, that’s the whole difference. Speed doesn’t help if the model doesn’t know what it’s looking at.

The private build question

There is something that goes beyond output quality, and it matters particularly to anyone working on something unreleased.

Public AI tools raise real questions about data. Where it goes, how long it’s kept, whether it quietly becomes part of what the model learns from next. For a product that hasn’t launched yet, none of that is abstract. Pre-release imagery sitting on public infrastructure is an exposure risk.

The alternative is dedicated hardware wired directly into production tools, on a closed network, running open source AI engines configured for the job at hand. Nothing routes out to a third party. Content doesn’t leave the building. For a sensitive product launch, if AI is going to be part of the process at all, this is really the only way it should be done.

The person running this on the day

None of it works without the right person holding it together.

A Post Supervisor sits at the intersection of production and post. On set, they are not there to observe. They are assessing what the shoot is producing against what post will need, identifying where AI can move a decision earlier, and coordinating whatever needs to be spun up to make that happen in real time.

If you have read the previous piece in this series, you already know the argument for having this person in the room. What AI adds is a new layer to what they can actually do while they are there.

The previz that stills production never had is now available. And it does not have to happen before the shoot.