Molten Cloud
Film Industry

AI in VFX: real use cases for film, TV, and games in 2026

How VFX teams use AI in 2026 for render denoising, textures, compositing, CG characters, game VFX, and generated shots, and the rights checks to run first.

Updated November 2026.

In 2024 Jory Federighi, a founding member of VFX studio Spruce, joined a Molten Cloud live talk on AI in visual effects. Clients kept asking Spruce, "Can't we just use AI?" Jory's answer was yes and no. AI could not make a sophisticated short film, such as a photoreal 2.5 minute scene of King Kong shaving his body bald. Jory added that anyone who has that video should send it over.

"We actually use very little generative AI creative solutions in our typical pipeline," Jory said. The value sat in tedious work that would be impossible, or highly impractical, to do by hand.

Two years on, both halves of that answer have moved. Generated shots now appear in finished streaming series. The tedious work has better tools, and the same ideas run in real time inside video games. This guide keeps the three cases, updates the tools, and adds what changed in film, TV, and games since.

AI in VFX at a glance

Task Tool or example in 2026 Section
Texture and material maps AI map generation from reference images 1
Background props at volume Image generators such as Midjourney 2
Render denoising Intel Open Image Denoise in Blender 3
Compositing with trained models Foundry Nuke CopyCat and BigCat 4
CG characters in live action Autodesk Flow Studio 5
Real time denoising in games NVIDIA DLSS Ray Reconstruction 6
Character deformation in games Unreal Engine ML Deformer 6
Fixes on footage already shot InterPositive, owned by Netflix 7
Generated final shots El Eternauta, Netflix (2025) 7

1. Textures and 3D materials

Spruce needed a "painterly town" texture for a green screen session for Spotify. Done by hand, that means sourcing or making materials, sometimes physically. Jory described it as literally watching paint dry. Spruce used AI tools to generate the full set of texture maps from specific references: diffuse, normal, roughness, and displacement. The team could iterate on materials and compositions in a fraction of the time.

Collage from Spruce's Spotify project: painterly reference images, a still of a guitarist on a crescent moon, and a paint texture made from reference photos beside one made with AI texture data

2. Background props by the thousand

For a video set in a record vault, Spruce needed thousands of prop records without showing real album art. The team used Midjourney to generate a unique cover for every record. It also made smaller details, like faded stickers on a milk crate.

Up close, the images show where they came from: a dog joined to another dog, a tandem saddle. At a glance they read as normal, and a background prop needs nothing more.

Stills from Spruce's record vault video beside a crate of records with AI generated album art and one example cover

The legal picture for image generators changed after the talk. In June 2025 Disney and Universal sued Midjourney in federal court in California, alleging it trained its models on their content without permission. Warner Bros. Discovery filed its own suit in September 2025, seeking up to $150,000 per infringed work. If you generate props today, record which tool made them and keep the prompts.

3. Render times cut by a factor of ten

The most useful example was the least glamorous. Intel's Open Image Denoise uses AI to remove noise from rendered frames in Blender. One still from the record vault took 9 minutes 45 seconds to render without it. With it, the same still took 55 seconds, about 10.6 times faster. Across a 490 frame sequence, that is about 80 hours of render time against 7.5 hours.

A side by side render comparison of the same 128 sample frame with and without AI denoising

The denoiser has since moved to the GPU. Blender 4.1 made Open Image Denoise GPU accelerated on supported NVIDIA, Intel, and Apple Silicon hardware. In Blender's own benchmark, the Junkshop scene denoised in 1.224 seconds on an Intel i9-13900K CPU and in 0.078 seconds on an NVIDIA RTX 3090. Full quality denoising now runs at interactive rates in the 3D viewport.

4. Compositing with models you train yourself

Foundry's Nuke has CopyCat, a plugin that trains a neural network on a few example frames. It learns the change from source to finished frame and applies it to the rest of the sequence. Foundry's own example is a garbage matte that has to hold across a whole shot.

Foundry added BigCat in Nuke 17.0, released on 26 February 2026. Foundry calls it "a new node optimized for large-scale dataset training." Artists can train on tens or hundreds of shots and solve across a full project.

5. CG characters in live action plates

In March 2025 Autodesk renamed Wonder Studio as Autodesk Flow Studio. It tracks actors in live action footage, swaps in CG characters, and matches the lighting. It also produces motion capture data, camera tracks, and clean plates with up to four actors removed. Results export to Blender, Maya, and Unreal Engine, or as USD for tools such as Houdini and Unity.

6. AI in game VFX

Games face the same render problem with far less time per frame. NVIDIA's DLSS Ray Reconstruction replaces "traditional hand-tuned denoisers" with an AI network. The network generates pixels in the parts of a ray traced frame where no rays were sampled. Version 4.5, announced in May 2026 for an August release, combines denoising and Super Resolution in one model. NVIDIA counts more than 1,000 RTX games and apps.

The second big use in games is character motion. Unreal Engine's ML Deformer trains a model to approximate high quality mesh deformation at runtime, including skin and muscle movement. Training takes a skeletal mesh, an animation covering a wide range of poses, and a geometry cache of target shapes. Epic recommends at least 3,000 training frames for an average character, and ideally 7,500 or more.

Both tools follow the rule Jory described. The artist designs the effect and the look. The model does the heavy computation that would otherwise not fit in the frame.

7. Generated shots in finished series

This is the half of the answer that moved most. In July 2025 Netflix co-CEO Ted Sarandos said El Eternauta used generative AI for a building collapse in Buenos Aires. It was finished 10 times faster than traditional VFX would have allowed. Sarandos said the cost otherwise "just wouldn't have been feasible for a show on that budget."

By July 2026 Netflix reported GenAI workflows in roughly 300 of its titles, adding that "in some cases, productions would have had to leave out key shots and sequences" without them. In March 2026 it bought InterPositive, Ben Affleck's AI tools company. Its model works on footage a production has already shot. It fixes continuity problems, missing shots, backgrounds, and poor lighting, and creates no AI actors. InterPositive trained it on a proprietary dataset filmed on its own soundstage.

Lionsgate went the same way with its library. In 2024 it partnered with Runway to build a custom video model from its film catalog. In June 2026 Lionsgate took an equity stake in Runway, and the two launched a joint program to develop new IP, starting with a short form episodic series.

What entertainment studios check before using AI VFX

The 2025 and 2026 deals point to the same checks. Run them before a generated shot goes into a deliverable.

  • What the model was trained on. InterPositive filmed its own data, and Lionsgate's model draws on its own catalog. A public image model carries the open questions in the Midjourney suits.
  • Who owns the output. Read the tool's terms before the shot is final.
  • Whose face and voice appear on screen. Disney's 2025 character licence with OpenAI excluded "any talent likenesses or voices." That deal ended in 2026. For the writers' side, see our post-strike AI playbook.
  • Which shots are generated. Keep a shot list with the tool and version used, so the answer is ready when a distributor or platform asks.

The last check is where VFX meets distribution. A film's AI history travels with it into every licence. Molten Cloud Rights Management lets distributors add indexed custom fields to titles and contracts. A field such as "generative AI used" then filters the catalog the same way a territory or a window does.

The 2024 rule still holds

Jory summed it up in the 2024 talk: "These tools are not as flashy or fun to play with but are making a huge impact on our daily workflow in VFX now." And on where generative AI belongs: "Where we really like to use generative AI [is] in situations where they're not making creative decisions for us or determining the composition. These are the things we like to do as humans. But they really take off the pressure and let us create things that are [otherwise] very time-consuming."

For more on AI across the film business, read the present state of AI in film and how streaming platforms use AI. Or watch the full presentation, Film and TV Operations in the Age of AI.