Foundry's CopyCat learns one effect from a handful of frames

The shot

Foundry's documentation for Nuke describes a node called CopyCat that, as retrieved on 16 September 2026, 'copies sequence-specific effects, such as garbage matting, beauty repairs, or deblurring, from a small number of frames in a sequence and then trains a network using machine learning to replicate this effect on a full sequence,' according to the documentation. A compositor feeds paired Input and Ground Truth images -- the example given is six garbage-matted frames paired with six hand-rotoscoped masks -- and CopyCat writes a trained network to a .cat file that a separate Inference node applies across the rest of the shot. The tool is restricted, per the same page, to NukeX and Nuke Studio. Foundry's own insights blog describes CopyCat's role on 'Dune: Part Two': compositors trained the node on hand-finished frames to remove tattoos, replace floor drains, and adjust Fremen characters' eyes, and the company states that forty per cent of the 1,000 eye shots completed required no additional touch-ups.

What the documents show

Both documents are Foundry's own. The documentation states what the node is built to do, learn one narrow, artist-defined transformation, and separates that from broader ambition: CopyCat 'is capable of generalizing to a wider variety of shots' only with 'significantly larger training datasets.' The blog account is a vendor case study, not an audited report; the forty-per-cent figure is Foundry's own characterization, attributed to the compositors, describing touch-up rates on one task, not a general accuracy claim. Together, the documents describe a small, per-shot model, trained and applied once, not a shared generative system.

The workflow

The practical shift is where training data comes from. Instead of querying an external generative model, an artist builds a small dataset from frames already corrected by hand, then lets CopyCat extend that correction across a sequence, feeding results back in when a shot needs further fixing. That keeps the effect specific to one artist's judgment about what 'right' looks like, and avoids provenance questions attached to models trained on external footage. A supervisor still decides which frames count as ground truth and reviews what Inference produces.

What the tool does not change

CopyCat does not replace the roto or paint artist's judgment about where a matte edge belongs; it automates propagating a decision the artist already made on a handful of frames. It is not a text-to-video or image-generation system, and Foundry's documentation does not describe it as one. This is editorial: a task-specific, per-shot trainer is a distinct category from a general-purpose generative model, even when both carry a 'machine learning' label.

• How many hand-corrected frames does a shot need before CopyCat's output is reliable enough to skip further review?
• Who signs off on the ground-truth frames a compositor chooses to train from?
• Does a vendor's case study report a general accuracy rate, or a result specific to one task on one film?

CopyCat marks how 'machine learning' gets used differently across a pipeline: not as a source of new imagery, but as a way of teaching a shot-specific tool to repeat what an artist has already shown it, once, by hand.