My Advice to New AI Creators

I must admit that advising new AI creators is a tricky balance for me to find. I got a lot of help on the way and I want to return the favor to the community. At the same time, I never asked for too many details from other AI artists that, I felt, was part of their closely guarded secrets.

And with the tremendous growth in AI and the difficulty it is to keep standing out, one can start becoming a bit protective of what one does to get the right magic. On the other hand, many of the things I do are very well known publicly.

So I don't intend to get into details of my techniques, but I'll cover the tools and services I use. That can be useful because many of you are now drowning in options and have no idea where to even start.

What do you want to make?

What you want to make decides a lot on what service you might want to pursue. If you would like to make SFW (suitable for work) content, your options are much richer and varied than if you want to make NSFW (not suitable for work) content. If you would like to make nice artistic non-pornographic images, then Midjourney or Lexica are probably the services I have had best experiences with.

But presumably you would not be getting advice from me if that is what you wanted. I make lots of naked girls with big tits and big butts. If you didn't care about that at all then you most likely wouldn't follow me or search for advice from me. So I assume you are going to want to make naughty images. In that case, your options are more limited, but that is what I will describe in the next section.

My Advice to Beginners

I would not recommend my current setup to beginners. What I would recommend however is Mage Space. That is where I really got going, making my best images. I still use it. I currently pay for a Pro Plan for $15 per month. But you might want to start with a basic plan of $8 per month.

It is easy to get going with. Offers a wide selection of Stable Diffusion AI models (checkpoints), thousands of LoRAs, Textual Inversions and private image generation. The latter is important. If you make naughty images you would rather not broadcast that to the whole world. If you don't know what a LoRA or Textual inversion is then don't worry. You don't have to know as a beginner. Just think of it as cool plugins which lets you get more fine grained control over the output images you make.

Mage Space also have a free service, but that will not let you create naught images.

Tools I Use

As described in an earlier post, I built a dedicated PC for AI Art creation, but I actually sit and do most of my work on an M1 Mac Mini.

Automatic1111

The main AI software I use is called Automatic1111 and has been installed on a PC running Ubuntu Linux. It runs as a web server which I connect to from my Mac. It looks like what you see below.

Of course, I don't just write a text prompt, hit generate and out comes a finished image I post.

Affinity Photo 2

Once I find an image I like it go through several steps to get a final good high-resolution image with good details. As part of that process I use Affinity Photo 2 for photo editing.

I cannot praise Affinity Photo enough. Unless you are a well-experienced professional, the Affinity tools will work better for you than what Adobe offers for a fraction of the price. I have owned and used Affinity Photo and Affinity Designer for many years now. Before I got into AI Art I used Affinity Designer to make illustrations which was used at my former work for UI mockups as well as in numerous tech articles. I had many people come ask me to use those illustrations.

I use Affinity Photo for manual inpainting. You can do that with e.g. Mage Space which has poor inpainting tools. Cut out sections of an image you want more details for and run that through your AI tool with the image-to-image generation feature. Next, you blend the result back into your original image. I typically use masks and eraser for such tasks.

Colors may not match, so you would want to use various tools for color correction and color matching. It is possible to add and remove cyan, magenta, and yellow from highlights, midtones and shadows in your image. I also adjust colors, contrast, vibrance, and color saturation for the whole image.

Topaz Gigapixel AI

Another important tool I use is Topaz Gigapixel AI. Here you have many options. Most AI tools come with their own upscalers. With Automatic1111 there are tons of upscalers to choose from. Upscalers are AI based tools as well. They take a lower resolution image and create a higher resolution image. Typically, upscalers are good at adding realistic texture. E.g. fur, hair, fabir and skin will look more detailed with an upscaler.

Other Services

An AI Artist will usually want to use a wide variety of services as not every tool will be equally good at everything. So here I want to mention some other services I use or have used in the past.

Leonardo AI

It has among the best interfaces I know for AI image creation. I would have highly recommended it if not for the fact that the censoring has gotten totally crazy. That is a general problem with online services today. They are righting the screws to avoid creation of whatever they consider offensive or inappropriate images. The problem is that this censoring has gotten so extreme that simply making normal images becomes challenging.

Lexica

Doesn't have a lot of options unlike Leonardo AI and Mage Space. But it this generator makes very crisp and colorful outputs. In particular I find it very good at making Robots and scenes with numerous characters in them or non-standard compositions. Unfortunately it is also a service taking censoring to hysterical levels making it very difficult to work with any image of a woman even if she is fully clothed.

Midjourney

I actually still use Midjourney a decent amount because it creates amazing looking images with quite simple text prompts. It is also a bit difficult with censoring but unlike Leonardo and Lexica I find it easier to understand the restrictions.

Midjourney, unlike Leonardo AI and Lexica, offers a ton of settings to tweak. But Midjourney has a number of downsides. While it is very hard to create ugly images with Midjourney it lacks stability in its output which I find makes it hard to refine outputs in multiple steps.

The other problem is that the interface to Midjourney is terrible. You create images by writing command to a chatbot on a Discord server. Yes, a very oddball interface choice indeed.

HuggingSpace

If you don't want to for out hundreds of dollars for a PC capable of generating AI images fast then an alternative is to simply rent powerful Nvidia graphics cards online to run AI services. There is a service called HuggingSpace which is a very generic service for running almost any machine learning software.

Hugging Space was my choice for running Automatic1111 software before I got my own AI Art PC built. The screenshot below is from HuggingSpace. Notice how you can select different graphics card. I was renting an Nvidia A100G graphics card highlighted in blue for $1.05 per hour.

Hugging Space is built on the idea of "Spaces". A space is your installation and configuration of some piece of software you want to run in the cloud which utilizes a graphics card. It is a community of users where people share spaces with each other. E.g. you can get started by taking a space configured for Automatic1111 by Camenduru and clone it. Once you have cloned a space you can modify it and run it on your rented hardware.

This approach is not very beginner friendly. You must be familiar with developer software stack. E.g. instance you have to know what a Docker container is. A space is defined by a Docker container. Within the space there is a Dockerfile which describes all the commands that will have to be run to install and create a docker container which will be able to run your Automatic1111 software in the cloud.

While this option works, I am not sure I will recommend it. For someone with my usage pattern it got too expensive in the end. I was spending close to $100 per month which didn't make economic sense long term.