Key points
- Artist-trained CopyCat models completed 40 per cent of approximately 1,000 Fremen eye shots without touch-ups.
- Training data comprised 280 existing shots from the first film, augmented to 30,000 eye images.
- Models were trained on artists' own data -- an explicit data provenance contrast with scraped datasets.
Summary
Foundry's case study documents the use of CopyCat machine learning models in the compositing pipeline for Dune: Part Two, released in March 2024. Artist-trained models handled approximately 40 per cent of around 1,000 Fremen eye shots without manual touch-ups, using a training set of 280 shots from the first film augmented to 30,000 images. The explicit use of the production team's own data as training material is a notable provenance element, directly relevant to teaching discussions about responsible ML pipeline design. The case study is already in circulation among compositing educators by word of mouth; this is the primary vendor account.
Related items
- DifFRelight: Diffusion-Based Facial Performance Relighting
- The Artist-Driven Innovation Behind the Films We Love
Source
Source: Foundry ↗ (Other)
Cite this item
Foundry (2024). ‘Foundry CopyCat ML on Dune: Part Two (Production Case Study)’, Foundry. Available at: https://www.foundry.com/insights/machine-learning/untapped-potential-ml-vfx
Your reference manager can also read this page directly: with the Zotero (or Mendeley) browser connector installed, save it straight to your library. Whole-collection exports: RIS, BibTeX, CSL-JSON.