M: I’m trying to do something for fun - are you aware of Rocky horror? - and I would like to learn how to deep fake well enough to create a fan video where the faces of the various characters are switched every scene. I have access to some video tutorials, but I’ve been burned in the past by thinking that I was going the right direction only to find out that I had accidentally stumbled down a very localized or specialized branch of knowledge, when I was looking for broader strokes. Can you suggest a starting point for this particular project?

G: For a project like this, you’ll want to focus on general deepfake techniques using software like:

  1. DeepFaceLab – The most widely used and customizable deepfake tool.
  2. FaceSwap – Open-source and a bit more user-friendly.
  3. ROOP – Good for quick face swaps with minimal training.

Suggested Approach:

  1. Understand the Basics – Watch general deepfake tutorials that explain dataset preparation, training, and blending techniques. Avoid jumping into niche tweaks until you have a functional pipeline.
  2. Collect and Prepare Data – Extract high-quality face datasets for each character (from Rocky Horror itself and other sources for variety).
  3. Train Models for Face Swapping – Train separate models for each character-to-character swap or use a generalized approach.
  4. Seamless Integration – Learn post-processing techniques (like color matching and edge blending) in software like After Effects or DaVinci Resolve.

Since you already have video editing knowledge, you should be able to handle the compositing side well. The biggest challenge will be training times and dataset quality.

Are you looking to automate the swaps as much as possible, or are you okay with manually processing each scene?

M: I can’t really answer that at the moment. I’m a functional autistic – I’ve done similar edits in the past, or accomplished similar projects, and sometimes I find a kind of satisfaction in doing it manually, while other times halfway through, I say, screw it and find a way to automate the task. So I really can’t say which will be the case in this circumstance. Why do you ask?

G: I asked because the choice affects your workflow. If you’re open to manual work, you might do scene-by-scene swaps with a simpler tool like ROOP. If you want to automate as much as possible, DeepFaceLab or FaceSwap with trained models would be better. Since you’re flexible, it’s probably best to start manually to get a feel for the results, then decide if automation is worth the effort.

M: Can you tell me if any of the options you suggested has a better publicly accessible, documentation, or any other advantages that would make it easier to learn? I don’t want to get halfway through and then realize that I’ve inadvertently joined a user group whose primary focus is making porn or whatever.

G: Good call. Here’s a breakdown of the options in terms of accessibility, documentation, and community focus:

  1. DeepFaceLab

    • Pros: Most widely used, lots of tutorials, very powerful.
    • Cons: Steeper learning curve, community has some questionable corners (but also legit users).
    • Documentation: Scattered but decent. You’ll mostly rely on YouTube tutorials and forum discussions.
    • Best for: If you want full control and the best quality, but expect some trial and error.

  2. FaceSwap (Open-source)

    • Pros: Transparent, decent GUI, community more research-focused.
    • Cons: Slower than DeepFaceLab for high-quality results, fewer tutorials.
    • Documentation: Actually has real docs (faceswap.dev) and an active Discord for legit discussions.
    • Best for: If you want open-source tools with a more academic/ethical community.

  3. ROOP

    • Pros: Very simple, real-time swapping, minimal training required.
    • Cons: Quality isn’t as good, limited flexibility.
    • Documentation: Not much, mostly GitHub discussions.
    • Best for: Quick tests or if you want a lightweight approach.

Recommendation:

If avoiding sketchy communities is a priority, FaceSwap is probably your best bet, since it’s open-source with a clear non-porn focus. DeepFaceLab is still the gold standard, but it requires more filtering of resources. ROOP is worth checking out if you just want to experiment fast.

Would you like me to dig up specific tutorials or setup guides?