Netflix's Eyeline Studios Says AI Gets Shots '60, 70%' There — The Rest Is the Fight

By Marcus Bennett

Netflix's Eyeline Studios has revealed it uses generative AI to replace locations, rebuild crowds and change lighting after filming — and it is building tools that could let directors reposition the camera once an actor's performance is captured. Mohsen Mousavi, the studio's VP of Virtual Pioneering, laid out the methods at Turin's View Conference on Tuesday.

Mousavi, whose credits include the final season of "Game of Thrones," drew a hard line between his studio's practice and image generation at large. "We're not prompting our way into shots and movies. That's not who we are," he said.

He calls the approach "hybrid production": live-action photography, conventional VFX and generative workflows combined.

How far does generative AI actually get you?

Mousavi put the honest number on the slide, so to speak. Generative models can carry a production roughly "60, 70%" of the way, he suggested, but "getting to the 100% is what matters the most." He was describing the gap between a persuasive AI image and footage that meets the consistency, resolution and creative-control demands of professional post-production — not a formal metric.

What does this change for a working production? The answer lies in Eyeline's "sparse set" workflows. Performers act on partial physical sets under relatively flat lighting. The surroundings and lighting are then rebuilt digitally, with the original performances retained.

"It's an extension of the traditional art department," Mousavi said, rather than a replacement. A generative art department explores designs alongside traditional artists before cameras roll.

A related volumetric-capture system allows camera perspectives and lighting to be reconsidered after filming. In demonstrations, Eyeline showed footage shot in one location and transformed into another — including environments built without green screens.

"It gives us both freedom of light and freedom of camera after shooting performances," Mousavi said, emphasizing the preservation of actors' identities and performances.

Where has Netflix already used it?

The techniques are not confined to the lab. Credits disclosed at the conference include:

• "Brazil '70: The Third Star" — the drama about the national team's 1970 World Cup victory. Eyeline built vast, emotionally responsive stadium crowds by combining filmed elements, sprites, conventional CG and generative techniques. Mousavi said the results "would have been very, very difficult to do within that budget" through traditional VFX methods.
• "The American Experiment" — the five-part historical documentary, for which he demonstrated background extensions, generative dust and other digitally constructed elements integrated with practical footage.

Eyeline also showed techniques for changing a performer's apparent age while retaining the underlying acting, and a research short exploring post-shoot camera and lighting adjustments.

What remains unsolved?

Plenty, by Mousavi's own account. Current generative fluid simulations are unreliable. "They're not deterministic, and they're very hard to control," he said. Eyeline wants artists to adjust water's velocity, viscosity and turbulence with keyframes, and to add foam or spray without altering the underlying simulation.

"Often in visual effects, what you really want to do is that you want to break the physics of the fluid," he explained, to achieve a specific creative effect.

Other demonstrations included restyling a Gaussian-splat 3D capture from winter into autumn, and AI-assisted relighting of human performances. One capture system uses 40 synchronized RED V-Raptor cameras to collect training data; Eyeline has also developed more portable equipment. Mousavi additionally showed research into constructing a navigable 4D scene from a single-camera recording.

He stressed that such experiments remain distinct from techniques already deployed on completed shows.

Built in-house, not bought off the shelf

Eyeline's systems rely heavily on proprietary work rather than off-the-shelf generators. "Most of the work we do is based on our own research and our own models," Mousavi said. The studio also uses outside partnerships and specially captured training data, though he did not identify its foundation-model providers.

The presentation lands as Netflix broadens its AI ambitions, including its acquisition of Ben Affleck's filmmaking technology company InterPositive. The streamer said in July that roughly 300 titles had already employed some generative AI.

For Mousavi, capability alone is insufficient. "We want to make sure that it's not a black box," he said, insisting that cinematographers and production designers must be able to shape the results. High dynamic range, 4K finishing and temporal consistency remain significant challenges.

"The goal is to make it completely cohesive," he concluded — regardless of whether a final image originated in camera, in traditional VFX or in generative technology.