What AI Changes in 3D Animation, Stage by Stage

When a brand asks for “3D animation” today, it can receive two very different things. One is a set of digital assets: a modeled product or a rigged mascot that can be posed, lit and rendered from any angle, as often as needed. The other is generated video that looks like a studio animated film and contains no 3D at all. Both are sold under the same name, and AI is involved in both.

In November 2023, DreamWorks co-founder Jeffrey Katzenberg told a Bloomberg panel that an animated feature that once took 500 artists five years would need less than 10% of that within three years, according to Cartoon Brew. That was a forecast, and its three-year window ends in November 2026. Whatever its verdict, AI has entered the 3D pipeline one stage at a time, and knowing where tells a brand what it is actually buying.

Six stages, each one a file you can reopen

A studio 3D film moves through a chain of departments. Modelers build shapes as meshes. Riggers give them a skeleton and controls. Layout places the virtual camera and blocks each shot. Animators perform the characters, pose by pose. Lighting and effects set the look, and rendering computes the final frames.

Rendering is where the numbers grow. GamesBeat reported in 2013 that each frame of Monsters University took 29 hours to render, on a farm of 2,000 computers with more than 24,000 cores, for over 100 million CPU hours in total. A befores & afters report on Elemental (2023) counts over 151,000 cores, against 294 for Toy Story.

That cost buys something specific: every stage leaves a file behind. If the client wants the mascot to wave with the other hand, an animator changes the animation while the model, rig and lights stay untouched. That separation is what a brand should protect when AI enters the chain, and it is exactly what generated video gives up.

Where machine learning already works inside the pipeline

The first place was the last stage. Path-traced rendering leaves grain in a frame unless it computes huge numbers of light samples. In a 2017 SIGGRAPH paper, researchers at Disney Research and Pixar trained a neural network on frames from Finding Dory to remove that noise, and showed results on Cars 3. Fewer samples plus a learned denoiser means fewer hours per frame.

Pixar went further on Elemental. Its fire characters were simulated as real volumes, then pushed toward images painted by Pixar artists with volumetric neural style transfer, run on a GPU farm bought for the job. Visual effects supervisor Sanjay Bakshi told befores & afters the result was “more organized and it’s more illustrative.” The style targets were the studio’s own paintings.

The 3D pipeline, and where AI assists

Six stages of a studio 3D production, with the machine learning work reported at each one

The 3D pipeline and where AI assists

Stage	AI assist
Modeling	Draft meshes generated from text or images
Rigging	Predicted skeleton and skin weights
Layout	Generated storyboards and rough blocking
Animation	Motion extracted from video; in-betweens filled from key poses
Lighting and effects	Style transfer on simulations
Rendering	Learned denoising with fewer samples
Generated 3D-look video	Prompt to frames, no mesh, rig or scene file

The middle stages are where research removes repetitive work. A 2020 SIGGRAPH paper from the University of Massachusetts Amherst predicted a skeleton and skin weights directly from a character mesh, the rigger’s first job. The same year, Ubisoft’s research group La Forge presented a network that fills in motion between an animator’s key poses, aimed at “lowering the workload of animators.” In May 2024, Autodesk bought a startup whose software extracts motion capture from ordinary video and applies it to a CG character.

Each of these tools hands its output back to an editable stage: a rig an animator can correct, keys that can be moved, a denoised frame from a scene that still exists.

Generating the model itself

Modeling is where generation reaches furthest. In 2022, researchers at Google Research and UC Berkeley showed a method that builds a 3D object from a text prompt using only a 2D image model, with no 3D training data, and exports it as a mesh for standard 3D software. Image-to-3D methods followed. The Animation Guild’s September 2024 report judged that 3D models could be generated from images and concept art “at a basic level.”

For a brand, the distance between basic and production-ready is the whole job. A generated mesh often arrives with messy geometry, fused parts and lighting baked into its textures, which make it hard to rig or relight. A product also has exact dimensions: a generated bottle can look right and still get the cap thread or the label curve wrong. The safer route starts from the manufacturer’s CAD data and keeps generated meshes for props, backgrounds and early drafts.

Video that looks 3D and isn’t

The other branch skips the pipeline. Asked for a glossy animated mascot, a video model produces frames that look rendered, with soft shadows and plush textures, but it never builds a mesh, a rig or a light. It predicts pixels, and nothing is left to reopen.

The speed is real. A concept clip, a mood test or a short social piece can exist in a day. But a new camera angle means a new generation, and the character can return with an extra button or a longer ear. Across separate shots the model must be fed reference images every time to hold the design, and the product in the mascot’s hand is redrawn in every frame.

The two branches can meet. A team can render the real 3D mascot or product in key poses and use those frames as references or first frames for generation, or composite a rendered product into a generated scene. The 3D asset anchors what must be exact, and generation brings speed to the rest.

What the animators signed

The people who run the traditional pipeline have written AI into their contract. In December 2024, members of The Animation Guild, the Hollywood animation union, ratified a 2024 to 2027 agreement with 76.1% in favor, Cartoon Brew reported. Its definition of generative AI explicitly includes 3D models, alongside text, video and images.

The terms leave studios room. A producer “may require employees to use any AI System” for covered work, but must give advance written notice when generative AI might be required, and consult the employee on request, including on alternative approaches. Prompts an employee writes cannot be used in a way that displaces a covered worker, and an arbitrator cannot bar AI use. Director Michael Rianda, quoted by Cartoon Brew, called the AI and outsourcing protections “not strong enough.” The guild’s 2024 report lists 3D modeling, rigging and animation among the most exposed crafts.

A brand hiring a 3D studio may therefore be buying work made with AI under rules it never sees. Asking which steps used it is a fair, ordinary question.

Choosing the right kind of 3D

Start from how long the asset has to live. A product that will appear in ads, e-commerce spins and packaging mockups for years earns a true 3D model, built from CAD where possible, because every future shot reuses it. A mascot that will star in campaigns and answer new scripts earns a model and a rig, and AI can speed up its rigging, in-betweening and rendering without giving up control. A one-off social clip or a pitch concept can be generated video, ideally guided by renders of the real asset.

Before signing, ask four things. Will you deliver the source files (mesh, rig, textures), and who owns them? Which parts were generated, and who cleaned them up? Can you render a new angle without regenerating the shot? Is the product geometry taken from our own data? A real 3D asset is reusable; a generated clip is a finished image sequence. Both have a place when the buyer knows which one is on the invoice.