An AI artist explains his workflow

How it works — and why it takes a surprisingly long time to make something good.

How does an AI artist maintain consistency with a recurring character? While AI art may appear to involve just a few clicks, it can be quite time-consuming.

The video above demonstrates the abridged workflow of the anonymous creator behind "Stelfie" — a time-traveling selfie-taker. The artist's process involves custom 3D-generated heads, initial sketches, and extensive toggling between Photoshop and the AI program Stable Diffusion to achieve the ideal appearance.

Throughout this process, the artist employs typical AI art tools such as inpainting (modifying specific image areas), outpainting (extending beyond the frame), and denoising (controlling image alterations). AI art can be as labor-intensive as traditional art, but its distinct outcomes make the final product unique.

Check out more of Stelfie here:

[MarkArandjus]

Nice to see a constructive use of AI image software being featured: not an automated process to replace the creative workflow, but a tool that creative people can use.

[-BONELESS-]

This is how I always envisioned what "AI art" is. Its not just about entering a prompt and saving the prettiest image of the bunch.. Using AI as a TOOL is what will make it that much powerful for artists

[mihirvagal]

"Now, all that I dream is about a bald 41-year-old dude. So no- not great" 💀

[mucc]

This proves to me that we should differentiate between "AI Art" and "AI Assisted Art".

[andrewcarr]

I have no issue at all with AI being used as a tool for artists, only when it is the artist and a human claims credit for it's creations. This is the best use case for AI in art.

[Westwdo]

The start needs a warning before people start having a seizure.

[dangmefinnish]

It's comforting to know that even AI struggles with hands.

[arothmanmusic]

I've been following this guy's work since he started. He is a master of the craft, which is saying something with such a brand new artform. Anyone can type in a prompt and get out an attractive image, but to produce something with real intention and creativity requires an artist.

[TheMIKESK]

This is the only instance I've seen of a workflow that really uses AI as a tool and not as a cheap way to generate an image and claim autorship. Unlike the majority of AI "artists" he doesn't conforms with whatever the AI produces with the promts, he really has an idea and a vision of what he wants and uses the AI to produce the different elements of the artwork.

[juliegolick]

So, what I'm seeing is, if you want a photorealistic image of real / specific people, there are probably better tools to use than AI. There are plenty of great pictures you could make with stable diffusion that would take WAY LESS than 17 hours and still look great, but photorealism with a very specific image and very specific person in mind is probably not one of them.

[Penultimeat]

This really shows that the major limit of AI image generation is actually just Composition. If we could immediately get the composition we’re looking for, then this process would be significantly trivialized

[kansSseiren]

This is a photobash in the form of text!

[fabianmosele]

it is probably a very different workflow for every person. Its nice to see that he also takes quite some time to create his images (cause I always thought those were purely generated). Sharing the workflow is key too, because it makes it very trasparent how one arrives to a specific image. I will share my short animated film done with stable diffusion and controlnet this friday, and I hope it will reinvent a bit the way one can think of these tools. Of course i will share a behind the scenes too, to show how meticolous the work was to create an entire short film (through real animation and ton of iterations). Thanks for creating all of these video related to generative AI!

[glitchtulsa]

This video seriously needs a seizure warning.

[SnowballAI]

Amazing workflow! Thank you for sharing.

[kaitlyn__L]

I thought it’d take longer to have recurring characters and themes, though it makes sense there’s a special network just for Stelfie’s face so it’s not quite as far as just being able to churn out images with your OCs.

This also cements my idea that no matter how good AI art gets, it’ll be at best just like commissioning a talented artist to finish your sketch. The main difference being how quickly it gets back to you with the next version after you ask for a change.

I appreciate that this guy is using his traditional art skills to direct the neural networks, too much of the comments are filled with people hoping it’ll be some amazing shortcut to save them from having to learn to draw. He recognises the fact that these systems can lead you down the garden path and lose sight of your original idea.

[TheMedicatedArtist]

I was confused until I remembered that with Photoshop, whatever image you have is final.

Like, the person who commissioned this piece likely doesn’t have pictures of themselves where he takes a punch to the face seconds after impact. Unless you’re a beast, you’re not gonna be able to draw that in Photoshop (yes, Photoshop is also a digital painting tool; literally anyone can do it).

Same with the boxer! The artist needs that exact pose to get the perspective right or else it throws the whole piece off.

[adamrobertscreative]

Now this is how you use AI…. Use it as a tool, don’t just add words and take directly what’s generated…. Awesome work

[MrKubahades]

Interesting. I make my art mostly with 3d software these days instead of paintibg by hand. This seems like useful tool, just the plagiarism is a huge problem

[IvynaJSpyder]

Dude could have just painted it himself with that time...