AI by department · Virtual production and real-time
This is the department where AI stopped being a separate conversation.
Machine learning has been inside the real-time pipeline for years: denoising a path-traced render and upscaling a frame are both learned processes, and without them a volume would not hit frame rate. The question is which parts of the environment pipeline it is good at, and which parts it quietly makes worse.
Last checked September 2026. Nothing on this page is legal advice.
What it does badly
What it does badly.
A generated backing has no true depth
Cuebric, from Seyhan Lee, has generated segmented 2.5D environments for volumes from a text description since February 2023. 2.5D is the limit and it is also the point: with no true depth, a generated backing holds for a locked-off or slow move and falls apart once the camera translates far enough for parallax to matter. That is a shot-design constraint to agree at prep, in the same conversation as lens and move.
Generated geometry is not real-time geometry
Image-to-3D output comes back dense, with no clean topology, no usable UVs and no levels of detail. A real-time artist rebuilds it.
Colour management is where it breaks
Generated content arrives unmanaged, usually in sRGB, into a pipeline built on a calibrated panel. It clips, shifts off-axis, and does not match the practical light on the actor’s face.
Nothing generative is deterministic
A volume needs the same frame at the same time on every panel, every take. Anything generative has to be baked in advance.
Generative tools have compressed the front half of environment work and barely touched the back half. Preparation, optimisation and colour are still the job.
What to learn now
What makes you more employable on a stage.
Real-time budgets
Draw calls, polygon counts, texture memory and shader cost. Being the person who can say an asset will not hold frame rate before the shoot day
Start here
Colour, properly
A managed colour workflow end to end, panel calibration, and how content should be delivered so it survives the wall
Start here
Camera tracking, lens data, genlock and timecode
The plumbing. It fails more often than the content does, and almost nobody on a crew can diagnose it
Start here
Where the line sits
Before you put generated content on a wall.
The exposure in virtual production is rarely creative. It is that an environment built in three days in 2026 turns out, three years later, to be unlicensable for a territory the distributor has just sold into.
| The question | Why it decides something |
|---|---|
| Which model produced each asset, not which interface? | Indemnity follows the model. An asset generated inside a major vendor’s tool using a third-party model may carry no indemnity at all. |
| Is the environment owned or licensed, and for how long? | Volume content gets reused across productions. A licence that covers one delivery does not cover a library. |
| Does the licence reach every delivery territory and term? | Territory-limited and term-limited asset licences are the most common way this becomes somebody’s problem later. |
| Were people or places captured? | A splat or scan of a location can contain identifiable people who consented to nothing. |
