Face Reference: Preparing Material for a Swap
The reference is what the model learns the face from. Its quality sets the ceiling on the result: improving the output face beyond what went in is impossible.
How many frames
You can work from a single frontal portrait, but every additional angle removes an area of guesswork. A sensible minimum is several frames from different sides: full face, two three-quarter views, a slight tilt up and down.
Which expressions are needed
A neutral face, a light smile with the mouth open, and closed eyes. The last is often forgotten, yet it is where the model learns what the eyelids look like — otherwise blinking comes out notional.
Technical requirements
- The same person, with no sharp changes of appearance between frames
- Even soft light: hard shadows get baked into the model and resurface in someone else's scene
- No glasses if the target video has none
- Sharpness on eyes and lips — these are the anchor zones
- No heavy retouching: smoothed skin deprives the model of texture
What cannot be fixed later
If the whole reference set contains no frame with a turned head, and the target video has the person turning away, no amount of parameter tuning will produce a good result. The model has simply never seen that part of the face.
The same goes for beards, moles and scars: absent from the reference, they will not appear in the result; present in it, they will appear everywhere, including frames where they should not.
Go to the site →Read next
Why Deepfakes Jitter on Turns, and How to Cure It Model Consent: Why There Is No Work Without It Neural Networks for Adult Content: What They Can Do Today How Image Generation Works: Diffusion in Plain TermsThis material is intended for adults over 18 and is informational in nature.