Introduction

One problem with SD images are faces and hands. Both are often deformed, malformed or low in details. If a given image is ugly with respect to faces one need a way to fix this problem.

What the Workflow Does

The workflow is fixing malformed faces in a multistep process. If a face is recognised as a face this will work, otherwise not. Input is a given image, output a modified image with fixed faces.  

Face Fixing

Faces are fixed on base of a given image. There is for the moment no need for changing parameters. It is proposed to used minimum a two pass.

I decided after some test to perform FaceDetailer as four pass. A preview shows the progress on each stage. Embeddings as well as LoRAs can be used to modify the output slightly.

One can leave positive and negative Prompt as it is. There is normally no need for a modification. Other settings works also out of the box.

Known Problems

It can be that the GPU VRAM runs out of memory while processing, if the input images are to large. I tested the workflow with images of a resolution of 2048 x 2048 pixel.

Upscaling

The upscaling module is for the moment bypassed. If one has to fix faces in an upscaled image, it makes no sense for upscaling to higher rates in resolution. t is propoe