What Types of Shot-Level Machine Learning Is Nuke CopyCat Suited For?

Nuke CopyCat is suited for repetitive, shot-level tasks with clear input-output relationships. Teams provide input frames and ground truth frames to train a model for Inference. .cat The trained model is then applied to the sequence. Its value lies in converting completed local repairs or marker removals into repeatable processes, though models, samples, and manual corrections must still be managed together.

First, confirm whether the task can be clearly defined.

CopyCat trains on correspondences between images. Local repairs, tracking marker removal, and fixed-region processing are all viable candidates. If the team cannot clearly articulate the differences between input and target images, preparing samples becomes difficult and results are hard to evaluate.

Ask three questions before starting.

Do the samples cover motion, lighting, and composition changes within the shot?
When results fail, can the issue be traced to specific image types or input conditions?
If automated results are unsatisfactory, can the project revert to traditional nodes and manual corrections?

Answering all three questions affirmatively makes investing training time safer. If the task requires frequent creative pivots or unique processing rules per shot, retaining traditional nodes is often easier to maintain.

Sample preparation matters more than parameters.

Input and target frames must correspond one-to-one, with consistent dimensions, cropping, and color grading. Samples should focus on the specific issue to minimize interference from irrelevant background changes. For shots with significant motion, samples must cover representative frames rather than just the cleanest one.

After preparing samples, retain the original inputs, manual corrections, and frame selection notes. This allows the team to review specific samples if the model later fails on certain motion types, instead of guessing the cause.

What to record during training

The CopyCat training interface includes parameters such as epoch, batch size, crop size, and checkpoints. These affect the training process and intermediate results, while GPU availability impacts iteration speed. Initial tests need not maximize all parameters; first confirm sample relationships and output formats, then adjust gradually to identify issues more easily.

Record the model file, sample version, parameters, training duration, and checkpoints for every session. Model names should reference the shot, task, and version to avoid mixing files intended for different purposes. .cat Checkpoints saved mid-training should also be labeled separately from final models.

Inference is another round of acceptance testing.

Inference applies the trained network to sequences. It requires rechecking input dimensions, cropping, color processing, and output channels. Any mismatch between training and inference settings may cause edge, color, or channel issues.

Before full processing, run a short sequence to observe motion changes, edge details, and visual continuity. If anomalies occur, determine whether they stem from input configuration, insufficient sample coverage, or unincorporated manual corrections before adjusting the model. Do not use full-sequence results to mask local errors.

Prepare for both delivery and fallback simultaneously.

After inference, archive the model, input and target samples, version info, and manual correction logs. Retain traditional node or manual correction branches in the project to ensure a clear fallback path for automated results. The team can compare processing time, correction counts, and final visuals for identical shots to decide whether to reuse the model elsewhere.

These ONCE proprietary product stills are shown here to illustrate the context of product visual content and do not represent CopyCat training or inference results.