CopyCat for Nuke Machine Learning
Taught by Doug Hogan
- Duration:
- 2 hours 54 minutes
- Software Version:
- 16
- Launch Date:
- February 2026
- Course Number:
- NUK253
vfx
AI
vfx
AI
In this project-based course, Doug Hogan walks you through building a complete ML-assisted roto and beauty workflow from the ground up. You will train a CopyCat model on a real hero shot, iterate on it through multiple refinement passes, extend it to a second shot, and explore creative and advanced use cases. Along the way, you will gain a clear understanding of how CopyCat works under the hood, how to design strong datasets, how to read validation loss and training curves, and how to recognize when a model is improving, stalling, or drifting.
A core focus of the course is practical, responsible use of machine learning in production. CopyCat’s core models are trained on licensed data, and your custom training only uses the images and ground truth you provide. This makes it a commercially safe tool for studio work, with predictable behavior and clear creative ownership. You will learn how to structure your training data, choose appropriate model sizes, batch sizes, and crop strategies, and decide when to retrain, fine-tune, or lock a model for production use.
This intermediate course is designed for artists who are already comfortable in Nuke and want to integrate ML into real-world workflows. Whether you work in roto, paint, or general compositing, this class shows how CopyCat functions as a practical problem solver rather than a black box effect.
Doug Hogan is a Creative Technologist and long-time Nuke compositor with experience across feature film, animation, theme parks, and commercial work. He teaches advanced Nuke courses at fxphd and works closely with studios and artists to integrate emerging tools into production pipelines. His approach emphasizes clarity, repeatability, and real production decision-making.
Join Doug in this hands-on course and learn how to confidently bring machine learning into your day-to-day compositing work.
Class Listing
Class 1
We begin by importing the hero shot and defining the roto and beauty problems we want CopyCat to solve. You will create the first clean ground truth frames and learn why CopyCat excels at patch-based problems. This class establishes the creative and technical goals for the entire project.
Class 2
We build the initial CopyCat graph and assemble the first training dataset. You will learn how to choose representative frames, prep ground truth, denoise the plate, and introduce synthetic variation where needed. We also cover cropped training strategies that help the model focus on difficult regions while maintaining full-frame context.
Class 3
This session focuses on the first training run. You will start the model, monitor progress, and evaluate predictions across the shot. We introduce validation loss, contact sheet previews, checkpoint evaluation, and how to recognize early warning signs of poor training. Core settings like model size, crop size, and batch size are explained in practical terms.
Class 4
Iteration is where CopyCat becomes powerful. You will analyze failure cases, add targeted ground truth frames, and retrain from the strongest checkpoint. We cover when to fine tune versus restart, how pretrained weights can accelerate training, and how to avoid overfitting.
Class 5
With a refined model in place, we scale it across the full hero shot. You will learn how to evaluate stability across motion and lighting changes, clean up edge cases, and finalize an ML-assisted roto or beauty pass in a production-friendly way.
Class 6
In this class we extend the model to a second related shot. You will test zero-shot performance, decide whether fine tuning is required, and use grading and preprocessing techniques to help the model generalize. This introduces multi-shot thinking and previews how future tools like BigCat will expand these workflows.
Class 7
We explore creative and advanced applications of CopyCat, including reflection and glint removal and subtle beauty adjustments. You will look at techniques inspired by other artists and see how the same workflow can support stylization, cleanup, or alternative looks beyond traditional roto tasks.