Currently works well with close-up photos of face, shots made from further away look bad - training data has only head shots.
Trigger word is HanniPham. I used 0.9-1.0 strength in my testing. If the face becomes oversaturated/contrasty then add "soft light, pale skin, smooth skin, soft lighting".
I used 512x512 imaged for training but 512x768 generations work well too. Use Img2Img for upscaling. Higher resolutions can cause twins.
This is work in progress model, I am planning to update the model with medium and further away shot in the future. Trained on ChilloutMix.