Improving “Hard-to-Edit-Facial-Images” in AI Renderings

Most people who generate AI images will, at one time or another, generate an image that is good except for the face.

John Iovine
5 min readDec 15, 2022

--

Images Rendered by author

Many times you can use the techniques I outlined in my article on Improving AI-Generated Faces.

https://john-iovine.medium.com/improving-ai-generated-faces-6561cf6ba255?sk=ec98ae77a709dc217a6db0757cbdfe02

GFPGAN and Baseten

Where Baseten and GFPGAN are weak is when trying to improve faces that are only a small portion of the image. Like with the image below. In this article, we will look at a few ways to work around that. Assume you wanted the following overall image. You would like to improve the face, as the eyes are kind of wonky.

Image with wonky eyes

Your first step may be to use the GFPGAN outlined in the above-mentioned article or Baseten website.

https://app.baseten.co/apps/QPp4nPE/operator_views/RqgOnqV

If you load the image in GFPGAN or Baseten the results are less than spectacular.

No significant improvement using BaseTen in this case

When those GFPGAN and Baseten tools fail, you can use the techniques described in this article.

Pixlr

One method that works, some of the time, is to copy the face portion of the original file to its own separate file. If you don’t own Adobe Photoshop or an equivalent photo-manipulative package, I suggest using an online tool Pixlr. This online photo software offers a free and paid version. For the work we are doing here, the free version will work for you. Go to Pixlr.com, and create a free account. Load the image in Pixlr.

Next, use the Marquee Select tool, and draw a box around the wonky face. Copy the selection by using Ctrl-C or go to file copy.

--

--

John Iovine

Science writer, thinker, self-experimenter, focusing on personal development and health — www.john-iovine.com