I’m a huge Stranger Things fan. The truth is, creating convincing close-up CGI faces is still very challenging today. The human brain is especially critical when it comes to faces, making them extremely difficult to “fool.”

After spending a lot of time working with DeepFaceLab in the past, I felt it was the right moment to explore the latest video diffusion approaches as a case study. As before, I collected a dedicated dataset for this project, but unlike DeepFaceLab—which is still extremely useful—I was able to add subtle details, such as small beads of sweat on the face, directly through prompting.

These video diffusion methods make it much easier to control nuanced, fine-grained enhancements compared to my earlier DeepFaceLab workflows.
Everything was created locally on my PC using open-source models, without using any cloud services.