Tutorials Motion

AI in VFX: Trained Networks, Relighting and Footage-to-CG

By Infinite Creation · Professional · ~15 min

A 1960s mid-century comic book style editorial illustration showing a compositor at a split screen where a plain landscape becomes the same shot lit by evening light, with three pinned reference frames joined by an arrow to a long film strip while a colleague checks with a clipboard.

Overview

VFX, short for visual effects, covers everything added to or changed in a shot after it was filmed. AI tools now do three jobs inside it. One trains a small neural network on a few hand-finished frames and repeats the fix across the shot. One relights footage after the event. One turns footage into an editable 3D scene. This guide is for compositors, VFX supervisors and independent filmmakers. It sets out what two makers state, what was unreachable for a third, and the checks that keep a generated result from slipping into a delivery unseen. We did not test any tool on this page. Each statement about a tool comes from the maker's own page as read on 8 October 2026, or from a trade source named in the Sources section.

What You Need

The untouched plate, the original footage of the shot, which you keep unchanged
A few frames fixed by hand to the standard you want repeated
Enough time to review every frame of the result, since an error often appears in only a few
A record of the tool, version and settings for the shot

Steps

Pick the job before the tool

Repeating a fix, relighting and building a 3D scene from footage are separate jobs with separate risks. A trained network inherits the quality of the frames you give it. Relighting changes how a shot looks. A footage-to-CG scene replaces a performer with a character. Name the job and the risk before you choose.

Understand how CopyCat works, from Foundry

A Foundry article says CopyCat, new in Nuke 13.0, lets an artist train a neural network specific to the shot at hand. It needs only a small set of before and after images, and the article's deblur example trained on eleven small crops. When training finishes, the network's weights are saved to a checkpoint, which the Inference node loads to apply the result across the sequence. The article names deblur, upscale and beauty fixes, says the uses are up to the artist, and does not say where training runs.

Understand what Beeble states about relighting

Beeble's page lists 3D Relight, which it describes as lighting a shot in full 3D, and SwitchX, which it describes as mask-guided relighting and replacement frame by frame. It lists BeebleCLOUD in a web app and Beeble Studio marked local. In its Enterprise section it says "We never train on your data" and offers a full commercial license. The page does not mention a Nuke plug-in. Our earlier notes cited a Beeble blog post about using it in Nuke, which we did not open, so confirm that with Beeble if your pipeline needs it.

Know what was unreachable about Flow Studio

Autodesk's own Flow Studio page returned an access error. A search summary of Autodesk's product pages, which we did not open, says Flow Studio is cloud-based, was formerly Wonder Studio, and turns footage into editable CG through markerless motion capture, camera tracking, masks and clean plates, with Maya, Blender, Unreal and USD export. Treat all of that as unconfirmed until you read Autodesk's pages.

Review every frame, not a sample

Step through the whole result at full size. A trained network fails on frames unlike its training set, such as a head turn or a motion blur it never saw. A relight fails at hair and glass. Mark every failed frame and fix it by hand or retrain with that frame added.

Keep the plate and log the shot

Keep the untouched plate beside the result. Record the tool, version, training frames and settings. A supervisor or a buyer will ask what was generated, and the record answers without anyone reconstructing it.

Pro Tips

Add the frames a network fails on to its training set and train again. That is the intended loop.
Do not train on footage you lack the rights to use. Read each tool's statement on training data.
Judge a relight on motion, not on a still. The flicker shows between frames.

Knowledge Base

What You'll Learn

In VFX the quality bar is set by the worst frame, so an AI tool is only as good as your review.

Why a trained network is a different kind of tool

Most AI tools arrive pre-trained and behave the same for everyone. CopyCat, as Foundry describes it, is trained by the artist on one shot, so the network knows only that shot. That makes it precise and narrow. It also means the artist is responsible for the training frames, and a fix that works on the sequence it was trained on is not guaranteed to work on the next.

Rights and training data

Beeble states it never trains on enterprise customers' data and that its plans carry a commercial license. Those statements are the maker's own and we did not read the licence terms. For any VFX tool ask where its models were trained and what the licence covers, because a studio or an insurer will ask you.

Where This Fits

A trained network or a relight changes a shot after it was filmed, and Blender for Motion Graphics covers building 3D elements and getting them into an edit.

FAQ

Q: Is CopyCat free?
A: The Foundry article we read does not say. It describes the feature inside Nuke, and we did not read pricing.

Q: Does Beeble work with Nuke?
A: The page we read does not say. Ask Beeble.

Q: Does AI replace rotoscoping?
A: The Foundry article does not mention rotoscoping. It names deblur, upscaling and beauty work. We have no evidence either way.

Vocabulary

Plate The original footage of a shot before any effects are added. Compositors keep it untouched so they return to it when an effect fails, and so a reviewer sees what was changed.

Inference node The part of Nuke, Foundry's compositing program, that applies a trained neural network to images. CopyCat trains the network and the Inference node runs it on the rest of a shot.

Relighting Changing the direction, colour or softness of the light in footage after it was shot. Doing it well needs the software to estimate the shape of the people and objects in the frame.

Clean plate A frame or shot of the set with the actors and unwanted objects absent, used to paint them out. Some tools generate one by filling the removed area from other frames.

Sources

Foundry's and Beeble's own pages were read on 8 October 2026. Autodesk's page was not reachable. Checked 8 October 2026.

Foundry, CopyCat guide. States CopyCat is new in Nuke 13.0, trains a network on the shot from a small set of before and after images, saves weights to a checkpoint the Inference node loads, and names deblur, upscale and beauty fixes. It does not say where training runs.
Beeble. Lists 3D Relight, SwitchX mask-guided relighting and replacement, BeebleCLOUD and Beeble Studio marked local, and the statement "We never train on your data" with a full commercial license. No Nuke plug-in is mentioned.

Autodesk's Flow Studio page returned HTTP 403. What this guide says about it comes from a search summary we did not open, and is flagged as unconfirmed. The Beeble Nuke plug-in claim in our earlier notes did not appear on its page, and the blog post it came from was not opened. No page states accuracy, failure cases or prices we relied on.