Tips from an AI art master

I got into a conversation on Mastodon with @[email protected], an account that posts absurdly good AI art, and I asked the innocent newbie question: how do you even do that? Improbably, the person took the trouble to really explain. The rest of this post is to record their answer for future use by me and maybe you.

This is the image (caption: “sisterhood”) that started the conversation:

I asked “As I scroll through your work, “I wonder about your setup. How do you approach the work to get such high quality results? Any advice to relative newbs like myself?”

The person replied as follows.

Do NOT bookmark – gone in 30 days!

Hoo boy! That’s a $64M question, with no simple answer. In a nutshell, it’s about learning how to assemble a good prompt, which will generate a good image, using a good upscaler, which not only makes your tiny image usable, but also adds lots more detail to it, and then using software tools to finish the image.

No way I can make this short, so I’ll do a post with multiple replies, and number them so you know what order to read them in.

PROMPTS

There are plenty of tutorials out there on what format to use for a good prompt, and though you can use the same prompt for the different AI’s, to get really good results you need to learn the nuances of each system.

Visit PromptHero to see what prompts other people are using, It’s a dying site, and you need to create an account, but it’s free (no hidden fees) and worth your time as a learning tool.

Visit the midlibrary website to learn about artistic styles. They’re really important. Though verified for use on Midjourney, these styles are also usable in some other AI’s, like Stable Diffusion.

UPSCALERS

There are no free good upscalers. The various AI’s have their own upscalers, and do a decent job, but to get really good upscaling, while adding incredible amounts of additional detail, you’ll need to pay for it.

My goto is Topaz Gigapixel. It’s no longer available, but they do have an all-in-one tool that combines the now defunct Gigapixel with two other discontinued software packages, which I also use. It’s Photo AI. You can get it for $150.

The new kid on the block is Magnific.ai. It’s pretty incredible, but it’s a ridiculously overpriced subscription service. If you got the money, go for it. Otherwise, get PhotoAI for a more cost-effective solution.

FINISHING SOFTWARE

The AI’s put out decent results. The upscalerrs make the output even better. Finishing software takes those results to the final level.

Any decent photo/image editor will enhance your image. There are free ones (the Gimp), non-free ones (Photoshop), and online services (don’t use them, so can’t name any).

I use Lightroom, Photoshop, Topaz Denoise, Topaz Sharpen, and Luminar Neo… on every Image I keep!

I load the images into Lightroom. It’s my hub for all I do, allowing me to manage my image assets and to make use of all my finishing software.

I tell Lightroom to use Photoshop to edit an image, using individual layers for each of the following steps.

I tell Photoshop to upscale the image 6x using Topaz Gigapixel — my 1344×896 Midjourney image becomes 8064×5376. Not only is the image upscaled, but the upscaler adds an incredible amount of extra detail.

I tell Photoshop to run Topaz Denoise to take care of any grainy issues.

I tell Photoshop to run Topaz Sharpen to add more detail.

I tell Photoshop to run Luminar Neo to do my final color/contrast/artistic touches.

All of the above is done using a Photoshop action script — it takes care of the drudgery stuff, and lets me focus on the creative stuff.

I save the photoshop file back into Lightroom.

AI PLATFORMS

There’s quite a selection to pick from.

Dall-E3 is best for being true to your prompt, but while good for illustration, it’s not that great for photorealism. That’s intentional, as are the other restrictions like using product names, the names of live authors, nudity, etc. etc. I like the illustrative results, but I don’t use it.

Stable Diffusion is for the hard-core crowd. Though the default model has restrictions similar to Dall-E3, it is open sourced and the community has created models that allow you to do anything you want. The caveat: You need to run it on your own machine, with a semi-powerful GPU, and a somewhat steep learning curve. Try the Night Cafe web site to play with Stable Diffusion. You can earn some free credits there that allow you to use it for free… sparingly.

For a newbie, I recommend Midjourney. It’s a subscription service… the cheap plan is under $9 a month, and you don’t need to worry about running out of GPU time — help train their AI by using the Rate Images option on their main web page and you’ll get a free hour of GPU time for every 30 minutes you spend with the AI. Unlike your subscription hours, the earned hours roll over until used up.

Midjourney will give you quality results with minimal work. It’s designed to output beautiful images, no matter what. Yes, it does have restrictions (nudity, gore, etc), but it’s nowhere as restrictive as Dall-E3.

Find images you like on the Midjourney website Explore page, copy their prompts, and try them out on the Midjourney discord page (and soon on their website). Play with the parameters like –s and –style to see how they affect results.

Learn how changes to your prompt affect the image, not only in the order of words, but in what’s said and not said, and how to weigh particular words to give them more importance.

Learn how to have private DM’s with the Midjourney bot, that way what you do is private and not viewable in public. But, do visit the public rooms to see what other people are doing and gain some inspiration for ideas you might want to try.