Faceswap Forum

Hey all, new to the forums. The future truly is now.

I saw in the guide, in training, your input B should have between 1,000 - 10,000 faces, ideally with varied expression.

Say I were to record a 60 FPS, 1080p video of a friend for face swaps. What would the best practice to obtain useable training faces be?

Sort of a niche question, but I'm going for indistinguishable from the original video level (once I figure out exactly what I'm doing)

Thanks guys!

Shoot the video in various locations/various lighting situations/various poses. Maybe just have him interacting naturally with stuff, as you need as full a range of expressions and angles as possible.

When it comes to extracting, extract every N should be somewhere between 30-60 (i.e. every half a second to a second)

I've done deepfake work for clients who shot their own footage, and my suggestions would be:

Aim for about 15-20 minutes of high-quality video footage of each person's face to obtain enough images to make a face set for each person. Medium distance, front face views, as in an interview, work best.

The background of the face should be a neutral color or darker than the face to be swapped. Avoid bright light behind or above the face, even lighting is best. Fast motion of the head and motion blur is problematic.

As torzdf said, the video of each person's face should contain a variety of face angles, lighting settings, and expressions. Think of the model as trying to re-create a 3D image of the face, or of the head moving 180 degrees from shoulder to shoulder.

Profiles or side views of a face can't be easily deepfaked. Extreme close-ups of the face should be avoided, because of the model's face resolution limitations. The face is typically trained at 128x128 to 256x256 pixels and that is a tiny portion of the whole video frame.

Thank you guys!

It's important to remember that there is a difference between Training data videos and the swap video. Making a quality training set needs a huge variety, but the swap video you want to minimize complexity in any way you can.

For training data, just run the camera, catch video when not actually filming the scenes. Everything you can get helps.

For swap, keep things simple, and straight forward.

Should I avoid flash?

Flash, probably, but you should definitely use set lights. Lighting is probably the most important thing that people miss out on.