Cattery shares finished models, CopyCat trains new ones

The shot

Foundry's own documentation describes Cattery, as retrieved on 16 September 2026, as 'a library of free, third-party machine learning models created using CopyCat' that 'bridges the gap between academia and production,' according to the current Nuke documentation. The page lists model categories, 'segmentation, depth estimation, optical flow, upscaling, denoising, and style transfer,' and describes access as downloading a model from a companion community site and copying its folder into a local Nuke configuration before it appears as a node. Foundry's own blog post introducing the feature, dated 6 December 2022, states Cattery arrived with Nuke 13 and was expanded in Nuke 13.1, and frames it plainly: 'Cattery doesn't require you to own any cats to use it,' only 'a free library of machine learning models that can run in Nuke that everyone has access to.'

What the documents show

Both documents are Foundry's own, and together describe a two-part system: CopyCat is the node an artist uses to train a new, shot-specific model, while Cattery is the distribution layer where a finished model can be published and downloaded by someone else. The blog post says the two are staged differently over time: 'Foundry is going to convert an initial set of models and upload them for artists to access but our long-term goal is to open the site up for user submissions,' meaning the 2022 library began Foundry-curated, with open submissions a stated future goal, not a shipped feature then. Neither document claims models are reviewed for accuracy beyond category labelling.

The workflow

For a compositor, Cattery changes where a model can come from: instead of training a new CopyCat network on a shot's own frames, an artist can download an existing model built for a general task and drop it in as a node. Foundry frames this as available regardless of studio size: 'it doesn't matter if you're a large studio or a solo artist, anyone can use the models hosted there.' A supervisor still judges whether a downloaded model performs well enough on a specific shot, or whether it needs a CopyCat model trained from scratch.

What the tool does not change

A model downloaded from Cattery was trained by someone else, on footage the downloading artist did not choose; that is a different provenance situation from a CopyCat model trained on an artist's own frames, even inside the same node system. This is editorial: free and community-sourced describe the distribution model, not a guarantee about a given model's suitability for new footage.

• Was a given Nuke machine-learning node trained in-house on this shot's own footage, or downloaded pre-trained from Cattery?
• Does a downloaded model's description name the task and data it was built for, and does that match the current shot?
• Which Nuke version introduced the specific Cattery or CopyCat feature a pipeline is relying on?

Cattery is best read as Foundry's answer to a limitation of CopyCat, that training a new model per shot is slow, by adding a library where a finished model can be reused, with the studio framing it as curated first and community-submitted second, not the reverse.