Filmmaking was never flat

A cinematographer walks onto a set and makes decisions with her body. She stands here, not there. She tilts the camera two degrees. She notices the actor's shoulder is blocking the practical lamp and asks the grip to flag it. These are spatial decisions made in three dimensions by a person occupying three dimensions. No part of that process has ever been flat.

AI filmmaking compressed it into one dimension. A line of text. A sentence describing depth, height, width, angle, and distance in a medium that has none of those properties. Then everyone wondered why the output felt like it was directed by someone who had read about cinematography but never stood behind a camera.

Autodesk Flow Studio, the company formerly known as Wonder Dynamics, launched a product on Tuesday that quietly inverts the entire paradigm. 3D Editor + Canvas gives the filmmaker a three-dimensional viewport over AI-generated imagery. You can rotate around characters, reposition them in space, adjust camera angles and focal depth, manipulate environments. The generation happens in 2D. The direction happens in 3D. Nikola Todorovic, the co-founder, put it plainly to The Hollywood Reporter: "We've seen a lot of these AI models get better, but they're really focused on 2D, and as filmmakers we want to control things in 3D."

This is the same Todorovic who stood at the Cannes Film Market in May and drew a clear line between pipeline AI tools and prompting-a-movie. He distinguished the two categories while most of the industry was still treating them as one conversation. Now he built the distinction into architecture. A filmmaker can drop an AI-generated character into an existing world and manipulate them with the spatial precision of a CG compositor, or take an existing character and shape an environment around them. What he calls "performance transfer" is the process of shooting an actor in an empty space and placing them into a generated environment afterward, controlling the result not through words but through a camera you can actually move.

The timing tells you something. Two days earlier, the EU AI Act's Article 50 took effect, requiring disclosure labels on synthetic content and offering an editorial exemption for work under human creative control. The same weekend, Puck's Matthew Belloni told WIRED that "everyone is doing it" and described an unnamed showrunner who castigated her writing room for not using ChatGPT. Netflix's 300-title disclosure sits on the table. Nolan's Odyssey crossed $911 million worldwide, zero AI, IMAX 70mm, a man who insists on the physical world earning more per weekend than most AI studios have raised in total.

In the middle of all that noise, a company dropped a tool that does not generate footage, does not write prompts, does not replace anyone. It restores spatial control to the person who should have had it the entire time.

The dimension that was missing

One hundred and fifty-one articles in this series have documented what the text prompt can and cannot carry across the gap between filmmaker and model. Composition. Lighting direction. Camera movement. Environment. Performance. Each article identified the same structural limitation: the filmmaker knows the decision in three dimensions and must compress it into one. The model receives one dimension and must hallucinate the other two.

That hallucination is where the defaults live. Center frame. Eye level. Medium distance. The statistical average of all compositions the training data contained. A filmmaker who types "low angle, close-up, subject positioned left of frame" is fighting the compression. Sometimes the model hears it. Often it does not. The result is a negotiation conducted in writing between a person who sees the shot and a system that reads about it.

The 3D editor does not fix the generation. It fixes the conversation. The filmmaker generates a scene through whatever prompt they bring, then lifts the result into a spatial environment where they can do what they have always done: stand somewhere and look through a lens. The generation model still carries its biases, its beauty filters, its training-data averages. But the filmmaker can now correct the spatial decisions that text could never reliably carry. Camera height. Subject placement. Depth relationship between foreground and background. The angle that makes a character feel small or powerful or watched.

Those are not prompt engineering problems. Those are filmmaking problems. They were always filmmaking problems. The text box was the wrong tool for solving them.

Two architectures, one split

This series documented a similar split four months ago when NVIDIA published Lyra 2, which generated navigable 3D worlds from a single image. The vocabulary split into two halves: what the world looks like (still verbal, still needs structured prompting) and how you move through it (spatial, returned to the filmmaker's hands). Autodesk's announcement arrives at the same split from a different direction. Lyra 2 generates the 3D world. Autodesk lifts existing 2D generations into 3D for manipulation. One builds the space. The other lets you direct inside it.

The implication is the same. Structured cinematographic vocabulary concentrates on the seed image, the reference, the prompt that describes what the world looks like. Spatial control returns to the filmmaker's body. CinePrompt's value concentrates on the half that remains verbal: the 1,457 controls that describe lens behavior, lighting direction, color science, atmosphere, material texture. The other half, the camera placement, the blocking, the framing, belongs back in the filmmaker's hands where it started.

The cosmetic surgery analogy

Todorovic told THR that the tool aims to "limit the differentiation from traditional filmmaking." That sentence reveals the current state of the industry as clearly as any earnings report. The goal is not to make AI filmmaking look like AI filmmaking. The goal is to make it undetectable. Belloni's WIRED interview this morning confirmed the same dynamic from the studio side: everyone uses it, nobody wants to be caught using it, and the conversation has shifted from "whether" to "who controls what comes next."

The EU AI Act arrived Saturday to regulate a practice the industry is already performing at scale with the discretion of a poker table. Three hundred Netflix titles. A showrunner berating writers for not using ChatGPT creatively. A $587 million acquisition of an AI company. And the most popular film on the planet was shot on celluloid by a man who crashed a real ship.

Autodesk's 3D editor does not resolve that tension. It sharpens it. A filmmaker who can manipulate AI-generated scenes in three dimensions produces output that is harder to distinguish from traditionally directed footage. The editorial exemption in Article 50 rewards human creative control. The 3D viewport is evidence of creative control. The filmmaker who repositions a camera, adjusts a character's blocking, and corrects the depth between two planes is exercising the same directorial authority the DGA contract protects. The tool produces better filmmaking and stronger regulatory standing simultaneously.

The question Belloni asked this morning, who controls what comes next, has nineteen institutional frameworks attempting an answer. The 3D viewport suggests the answer might be simpler than the frameworks. The person who stands behind the camera controls what comes next. Whether the camera is physical glass on a physical dolly or a virtual lens in a 3D editor, the authority lives in the person who decides where it points and why.

Filmmaking was never flat. The text box made it flat for two years. The dimension is coming back. What the filmmaker carries into it has not changed since a person first picked up a camera and chose to stand here instead of there.

Bruce Belafonte is an AI filmmaker at Light Owl. He has never described a camera angle in fewer than three spatial coordinates and considers the minimum insufficient.