Viridian's AI Experiments

Artificial Intelligence is here! Really! Anything created with AI assistance, including stories, should be posted here.
Viridian
Posts: 1582
Joined: Wed Apr 15, 2009 10:03 am

Viridian's AI Experiments

Postby Viridian » Wed Dec 07, 2022 1:54 pm

As an artist, AI-generated art is a bit iffy. A great deal of respect goes to everyone who creates their own work, whether it's drawn or modelled, and AI-generated artwork cheapens the value of art. However, it IS a thing, whether we like it or not, and the advances in the algorithms are mind-blowing. After seeing some very good examples that are getting harder to distinguish from real art, I'm getting quite enthused. At the moment I am experimenting with AI work and seeing how I can incorporate it into my work.

Currently I am testing out NovelAI. It's mostly a story-generating AI, but it does have a very good image-to-image algorithm. What makes the platform challenging is using the right tags to generate the best results. Things like appearance, clothing, backgrounds, poses, and so on must be tailored to see if you can roll the most desirable image. Each generation costs credits (which are paid for). There's a whole community around figuring out the best tags to use.

The results for our quicksand fetish are a bit more obscure, but I'm kinda surprised. Since I'm not using most of these images, but as I'm technically paying for them, I figured I'd share both my results and my methodology for those who want to see for themselves.

The first set uses the "realistic" tag focusing on "park ranger". Using the "Full" mode, you can generate more NSFW elements, such as using "torn shirt" to show breasts and nipples. There are a bunch more tags you can use to fine-tune the results, but that's a deeper (ha) discussion. Some of these are from plain text tag generation; a few used one of my artworks as a source image. The difference I gained was the pose - I've attached the reference artwork. Since my piece was hip-deep, the AI more accurately replicated the hand position. It doesn't recognise depth too well, hence just supplying a generic model doesn't get good results, but a partially submerged model might do better.
You do not have the required permissions to view the files attached to this post.
Viridian @ deviantART: http://viridianqs.deviantart.com/

Viridian
Posts: 1582
Joined: Wed Apr 15, 2009 10:03 am

Re: Viridian's AI Experiments

Postby Viridian » Wed Dec 07, 2022 1:56 pm

A couple more. As I said above, the AI doesn't do a great job of recognising the intent of the artwork, and there's a lot of "uncanny" things, so better source material and more defined tags refine the results greatly.
You do not have the required permissions to view the files attached to this post.
Viridian @ deviantART: http://viridianqs.deviantart.com/

Viridian
Posts: 1582
Joined: Wed Apr 15, 2009 10:03 am

Re: Viridian's AI Experiments

Postby Viridian » Wed Dec 07, 2022 1:58 pm

This was my first experiment, before I learned about the "realistic" tag. The platform is anime-based, so it defaults to this style. Source image provided.
You do not have the required permissions to view the files attached to this post.
Viridian @ deviantART: http://viridianqs.deviantart.com/

Viridian
Posts: 1582
Joined: Wed Apr 15, 2009 10:03 am

Re: Viridian's AI Experiments

Postby Viridian » Wed Dec 07, 2022 2:11 pm

This set uses my "Gail's Gallop" illustration. I tried to generate the horse with the image, but it defaulted to Gail riding the horse even with the source image. To help the AI, I specified "red turtleneck sweater" and "mud" and "quicksand". The AI doesn't recognise mud and quicksand well, so the source images do a LOT to get the image right. I had to specify the posture more specifically to match my image (leaning back) as well as refining hair length and colour.
You do not have the required permissions to view the files attached to this post.
Viridian @ deviantART: http://viridianqs.deviantart.com/

Viridian
Posts: 1582
Joined: Wed Apr 15, 2009 10:03 am

Re: Viridian's AI Experiments

Postby Viridian » Wed Dec 07, 2022 3:02 pm

Jungle explorer theme, based on an old drawing. Again, the quicksand effects are somewhat impressive, but it really needs a good source image to get the right textures and setting, which makes a lot of difference for us. With proper editing and texture manipulation, these can be turned into very good final products, though I'm only interested in testing what they generate in terms of poses, outfits and settings. I had to specify the colour of the shirt.

It's worth noting that as an anime art generator, the breast proportions have to toned down. These are, FYI, "medium breasts". Large breasts are, well, anime proportions. Sometimes you do want that effect, but it gets ridiculous when you start showing more skin. But that IS a thing I know people do l like.
You do not have the required permissions to view the files attached to this post.
Viridian @ deviantART: http://viridianqs.deviantart.com/

duuudeization
Posts: 644
Joined: Mon Aug 06, 2012 6:43 pm
Location: Florida

Re: Viridian's AI Experiments

Postby duuudeization » Wed Dec 07, 2022 3:12 pm

Actually pretty good. Hats seem to throw it off, and body parts proportions ( arms too short, heads are small...not fussing at all about the breastesses and super thick thighs )...very intriguing. I appreciate you sharing all this
Its all about peace and love and good happiness stuff

Viridian
Posts: 1582
Joined: Wed Apr 15, 2009 10:03 am

Re: Viridian's AI Experiments

Postby Viridian » Wed Dec 07, 2022 3:28 pm

The AI does begin to learn and adapt to what you output. I've run quite a few tests (many that weren't very good), but supply it with source artwork. I didn't use my tags for this, mostly focusing on "3girls", "bra", "panties" and "night".

It gets more difficult to specify mixed items, such as one being topless, one covering breasts, etc. Adding tags like "topless" or "nipples" will mix and match. However, get too fancy and you generate some absolute monstrosities (not included, thank me).

What I'm seeing is that it really tries to stick to the source image, especially if you have the Scale on high (this narrows down how much freedom the AI has to interpret the prompt). Since I'm deliberately trying to recreate my source images, I'm keeping it high. The benefit is that it really gets the quicksand somewhat right, whereas without a source image the AI simply does not know what quicksand is.
You do not have the required permissions to view the files attached to this post.
Viridian @ deviantART: http://viridianqs.deviantart.com/

Viridian
Posts: 1582
Joined: Wed Apr 15, 2009 10:03 am

Re: Viridian's AI Experiments

Postby Viridian » Wed Dec 07, 2022 3:35 pm

Again, challenging to hit separate individual features with multiple characters.
You do not have the required permissions to view the files attached to this post.
Viridian @ deviantART: http://viridianqs.deviantart.com/

Viridian
Posts: 1582
Joined: Wed Apr 15, 2009 10:03 am

Re: Viridian's AI Experiments

Postby Viridian » Wed Dec 07, 2022 3:54 pm

And throwing in Quicky Sanders. I mean, the AI really tries, but you can see where it fall short. I tried writing it as prose, but it can't visualise exactly what I'm visualising. The source image helps a lot, but you have to add keywords to include more items and complexity. However, the AI again interprets the most common usage of these keywords. In most cases, having a "phone" means that the character is using the phone, and I can't get the phone to be on the ground or in the mud, since these concepts are not relevant to the AI. In trying to do so, it generate half a dozen phones on the ground with the character holding one. It was also baffling that the colour of the shirt wasn't consistent, the skirt was on or off, and the bra was always different in colour despite being "purple".

Here's the best one. If there's any solace to artists, the AI can't reach this level of artistic intent. Looks like it's really good for simpler items that are more open to what we want, but not very good for very specific things.
You do not have the required permissions to view the files attached to this post.
Viridian @ deviantART: http://viridianqs.deviantart.com/

Viridian
Posts: 1582
Joined: Wed Apr 15, 2009 10:03 am

Re: Viridian's AI Experiments

Postby Viridian » Wed Dec 07, 2022 4:18 pm

Here's something interesting. In this case, the AI replicated the composition of my illustration as the source, specifying "ancient ruins" with "mud" and "quicksand" along with specifics like "khaki buttoned shirt", "red long hair" and "cleavage". With the source image provided, it couldn't generate a quicksand image, although it did generate a very good portrait of the character.

However, when I removed the source image, it did create a few very nice poses. It's not what I wanted, but it's still pretty good. (EDIT: Actually, I accidentally reset my settings. Making the AI match the source image to the max created the 4th image, which is really close to what I had).

Bear in mind, there's a lot of re-rolling (read: gambling) with generating AI art.
You do not have the required permissions to view the files attached to this post.
Viridian @ deviantART: http://viridianqs.deviantart.com/


Return to “AI Creations”

Who is online

Users browsing this forum: No registered users and 1 guest