Fix Faces Usage Guide
What is the Fix Faces tool?
Fix Faces takes an existing creation, detects faces in the image, then creates an automatic inpainting mask for each face, and does an inpainting generation for each mask. The inpainting generation is zoomed-in to fit the mask area so that the face can be re-generated in high fidelity.
General advice
Fix Faces does a great job at fixing faces that haven't come out nicely in the first generation, which usually happens because the subject is far away (i.e. it's not a close-up portrait). It generally won't make an already good face look better, but it can help with faces that are blurry, distorted or deformed.
- Outputs will generally always match the input size but there are limits. e.g An upscaled creation will be downscaled to match the original image size to pass to the image-to-image model.
- Each detected face results in an entire image generation to regenerate that face, so credit costs are based on the number of faces.
- Please beware that if the face in the image is important (e.g. it's an image of a real person), Fix Faces will most probably make it unrecognizable and you should avoid using it. The exception is celebrities - as long as you include the celebrity's name, the Fix Faces will most likely improve their face.
- Your prompt should describe the face you want more so than the image you're enhancing. Write prompts like you would for Stable Diffusion inpainting.
- Negative prompts help take parts of the face out you don't like, so make sure to use them.