This technical research report explores whether deepfake technology can be effectively used in the visual effects (VFX) industry as an alternative to traditional CGI digital humans. It examines the underlying machine learning techniques, particularly auto-encoder neural networks, and evaluates their strengths, limitations, and practical applications in a production environment.
Through hands-on experimentation using DeepFaceLab, the report analyses key technical challenges such as resolution constraints, training data requirements, lighting inconsistencies, and warping artifacts. A public survey and real-world test cases are used to assess how convincingly deepfakes replicate human faces across different shot types.
The findings show that while deepfakes currently fall short of replacing high-end CGI, especially in close-up shots, they offer viable use cases in mid-to-wide shots, stunt doubles, and low-budget productions. The report also addresses ethical considerations, concluding that deepfake use in VFX can be acceptable when proper consent is obtained.
Overall, this research provides a grounded, practical evaluation of deepfakes in modern VFX pipelines, highlighting where the technology works today and where future advancements are needed.