The generator keeps whatever is in the two photos. Give it two clear faces and the clip looks like you both; give it a dark group shot and it guesses who is who.
One person per photo
Photo 1 is the survivor, Photo 2 is the one who turns. Each photo should hold exactly one person. You do not need a photo of the two of you together; two separate selfies are better.
Light
Even light on the face. Window light in daytime works. Avoid a bright lamp behind you, sunglasses, hats and masks.
Framing
Head and shoulders, or half-body, with a little space above the hair. Close crops lose the hair; wide shots lose the face.
Matching quality
If one photo is sharp and the other is blurry or tiny, the weaker face drifts first, usually after the cut into the memories. Use two phone camera photos of similar quality, at least 800 px on the short side.