AI art and retouching

Let’s talk about AI-art

The Midjourney and Stable Diffusion neural networks are very popular right now. The first creates beauty very simply, but is limited in functionality, the second is much more flexible, but in some ways similar to programming. In addition to prompt (query) you need to specify a lot of parameters, choose a model on the basis of which the art will be generated, adjust the number of steps, denoise, you can set the character’s pose, etc.

But what if you want to make art based on a photograph of a person? And not just so that it slightly resembles the original which is very easy to do now, but so that there was an exact recognizability of the person, but to replace both the body and the background. This too can be done. If you think it’s done in a couple of clicks — it’s not. Let’s see how much time it takes to make an art that looks like a person and how many images need to generate to get the desired result. I will use Stable Diffusion.

1 — First we need an idea. If it exists, we describe it and begin to select a model on the basis of which the images will be created. Depending on the model, the result can be drastically different. This can be photorealism, but it is not so interesting, or cartoon-style, can be like 3d illustration, etc. Slightly changing the prompt and changing the settings, I select the model. When the model is selected, then move on.

Our idea is Superwoman in the kitchen (It was a print for my friend Natalia for her birthday, just for fun).

2 — Now need to generate different images with different settings, which in the future we will combine and refine in Photoshop. Selecting a suitable image, we fix it and start creating variations.

3 — Now take the original photo and the generated images and go to combining it all in Photoshop. At this point and get the magic to recognize the person.

4 — This point could be the third, it does not matter. Refine individual parts of photos that are not very good in general art. For example his hands, fabric with a logo, or any other parts we need. And also replace it in Photoshop.

The highlighted red is the model selection.
Green is the main generation.
Purple is detailing.

That’s it. Now we get art with human recognition. As you can see, this is not just a 1 or 2 clicks on the neural network, but a whole lot of work. In this case, it took about 4 hours. Some could be done even faster, it all depends on the requirements and the picture itself. Yes, it is not perfect, it could still be improved some points, but for my task it was enough.

Ps — this is not the only way, when you can make a recognizable face. It is also possible to train a neural network on your face, gathering some amount of photos and getting an addition to the neural network and generating an image with you when needed. This method has its pros and cons. I have now described only one way to get the result you need.

And here are some more examples of different complexity, both simple art and more complex art.

If you need such an art, feel free to contact me, I’ll be glad to help you.
You can use any of my contacts: https://sergretouch.com/contact/