AI world models will be used as film sets. Throw in robotics and "let's generate a game" and they're rapidly emerging as the "spatial" version of more traditional LLM-based AI.

In AI filmmaking, coherence has been the holy grail.

Environments shift between takes. Lighting drifts. Spatial logic collapses. The result is footage that looks like a sequence of unrelated images rather than scenes from a single world.

World Labs' Marble platform is now being used by working filmmakers to address exactly that problem, and the approach is fundamentally different from prompt-based tools.

A Virtual Backlot, Not a Generator

Tim Simmons, a former Hollywood documentarian who runs the Theoretically Media YouTube channel (170,000+ subscribers), used Marble to produce Alarm, a 48-second micro-short built entirely around a single persistent 3D environment.

Rather than prompting individual shots and hoping for visual continuity, he treated Marble's Space Station world as a digital backlot: a stable, explorable 3D space he could move through, frame from any angle, and return to repeatedly.

"Marble gave me a stable 360-degree virtual set I could wander," Simmons said. "I kept finding angles I'd never think to prompt for without an actual space to explore."

His workflow threaded several tools through that stable core. Characters were built in Midjourney using 360° pose references and composited into stills exported from Marble scenes. Motion and timing were refined in Veo-3. Final assembly happened in Premiere Pro, with audio from Suno and ElevenLabs. Because the environment never changed, every iteration concentrated on performance and pacing.

Episodic Worldbuilding at Scale

Game animator Rik Vasquez, whose credits span Baldur's Gate, Star Wars: Knights of the Old Republic, and Jade Empire, pushed the approach further.

His project Cryptid Worlds used Marble Plus to produce twelve short episodes set within a single connected world, constructed by stitching four generated scenes into one .spz environment. That modular virtual stage then hosted multiple storylines, with scenes composed using Runway, Pika, and Sora layered on top.

"Being able to walk through a generated world and pick my own camera angles was a game changer," Vasquez said. "It felt like location scouting in an alien world, but fast and completely controllable."

How Marble Works

Marble combines AI synthesis with Gaussian Splat rendering. The resulting worlds are persistent: they can be reused, reframed, and extended across multiple shoots or episodes. That persistence is what separates it from conventional text-to-video pipelines, where each generation is essentially stateless.

Why This Matters

These are the first clearly documented cases of a 3D world model being operationalised specifically as a narrative continuity tool in film production. The consistency problem has been cited repeatedly as the blocking issue for serious AI filmmaking.

Marble's approach doesn't solve it by improving generation quality shot-by-shot; it solves it architecturally, by making the world itself the persistent asset rather than any individual frame.

Both Simmons and Vasquez describe plans to extend the approach: Simmons into camera movement and interactive character blocking, Vasquez into animated props, dynamic lighting, and branching storylines.

The shift they're describing, from prompting scenes to inhabiting worlds, may define how AI-native film production matures over the next production cycle; or at least points to a new toolkit in the production pipeline.

Case study published by World Labs on 12 November 2025.