Mr Wisp
patreon
The Rest: January 2023
🕑 Added 2023-01-01 23:45:00 +0000 UTCHello! Before you proceed, please take the time to note the FULL FOLDER OF 300 AI-GENERATED IMAGES ATTACHED TO THE END OF THIS POST! HUZZAH!
You may now proceed.
Notes on the AI Image Generation Process:
A quick primer for those unfamiliar: NovelAI is a paid service in which you spend virtual currency (Anlas) to have an AI generate prose or images for you based on data you provide it. I can't speak for the prose side of things since I have focused entirely on image generation, and specifically image to image generation (in which you upload an image of your choice to serve as the basis for what the AI generates) at that, but as far as that variety of image generation is concerned, here's what you need to know:
1. Tags or prose prompts determine nearly everything. Beyond the general details of the image you upload as a base, the image you generate is mostly limited by what you can communicate to the AI, which is conveyed primarily through the tags you use (You can also try writing sentences to convey the same information, with varying degrees of reliability). A small amount of tags will result in a fairly generic looking image, but too many tags might confuse the AI and lead to inconsistent results. Finding the happy medium for the sorts of images you want to generate takes up most of the time you'll spend with NovelAI's image generator.
2. Settings can help with quality and/or creativity. You can adjust the utility's various sliders to tell the AI to generate images based on numerous different qualifiers, such as the creative license it is allowed to take relative to the tags you've put in, the amount of noise (image detail) that will go into the image, how many variations the AI generates at once, how much you might want to enhance (or change) existing output, and how consistent you want the output to be relative to itself (accomplished by specifying a seed for the AI to use). Understanding and taking advantage of the little quirks of each of these settings can help you maneuver the AI's output toward more accurate (or exciting and provocative) results, but learning these quirks in the first place is a very overwhelming and always ongoing process.
3. Its specialty is anime waifus.
No, I'm fucking serious.
NovelAI has a few modes you can try out, but unless you want to try your hand at generating some fairly janky furry art, you're best sticking with the anime diffusion methods, which are based on data provided by the image databases of Japanese 'booru' image sites (danbooru, etc.) which are most frequently full of anime style images. Anime girls, specifically (though I hear you can generate a mean femboy with the right tag input). So measure your expectations accordingly. You can tell the AI to generate images in a certain non-anime art style, with varying degrees of success, but it tends to benefit the quality of your output just to go with the flow and accept that, if you're generating images of characters, they are now anime girls. C'est la vie.
Luckily, even with this latter limitation, there is still plenty of room to experiment and push the boundaries of the AI, which is always learning based on what people tell it to do. And experimentation is really the main draw here, which leads me to the purpose of this gallery.
The title Re-AImagined represents my intentions in creating these galleries: to reimagine and represent my characters in ways I would never be able to just by drawing them via my own methods. To clarify, this isn't at all me giving up art to focus on AI image generation, but it is allowing me to think about and develop my characters in different ways when I don't have it in me to sit down and draw it all out (something which is very valuable to me with my current 'work' environment not being very conducive to the kind of sustained focus I need to draw consistently.) The practical side of this venture is certainly there, but I wouldn't be doing it at all if there weren't substantial creative reward there as well.
AI image generation first caught my eye and has intrigued me since late summer 2022 due to the creative possibilities it opens up: especially the surprise results created by the gaps between what the creator communicates to the AI and what the AI ultimately produces. The results of these gaps -the images generated- admittedly fall on a spectrum ranging from extremely frustrating (as I hinted at earlier, learning the AI's behaviours, biases and blind spots is a herculean and usually vexing task) and occasionally cursed on a profoundly eldritch level, to truly breathtaking, and you will see examples of both ends of this spectrum in the attached folder.
Suffice to say though that it can take a lot of work to generate a reasonably consistent lineup of images, depending on the baseline of quality you're expecting. In my case, my goal every time I start a NovelAI session is to see how close I can get to generating AI images that are representative of my various characters without sacrificing a) details that are vital to the identity of those characters, and b) overall image quality. Anyone who has spent any amount of time fooling around with NovelAI, especially image to image generation, knows this is easier said than done. NovelAI is a fantastic asset for generating fairly generic anime-esque settings, characters, and objects, but the more specific and ambitious you get with the details you input, the more likely NovelAI is to misinterpret or miss something. That said, the AI has its clear tendencies regarding what it knows and doesn't know how to do, most clearly represented by some tags having greater influence and/or appearing more consistently across batches using the same set of tags. There are also many small tricks and techniques you can learn to help the AI better grasp your meaning, which is part of the fun of learning how to use the utility.
This month's gallery is comprised of the results of three image generating sessions from varying points in my NovelAI experience. The earliest of these sessions was the Mallory session in early November 2022, followed by the Natasha session in late November of that same year, then the Leslie session in mid-December. I'll discuss the three sessions in order, providing tips and anecdotes to do with the image generation process when able.
Mallory
This cluster of images turned out exceedingly well despite (or because of) my relative lack of experience with NovelAI at the time. By that I mean that I played it safe with this batch and spent a lot of time with the various settings kept very close to their defaults. For instance, when you first upload an image to serve as your base, the AI's Strength slider is set to 0.7, meaning the image you generate will be unique and fairly fleshed out, but still clearly based on the base image. The early Mallory images in this set exemplify this output: like with all my image to image generations, I use an existing image of the character I want to represent as the base (usually their reference image). Hence why Mallory consistently has one leg folded in at the knee and the other outstretched. To get more varied and interesting results, you'll usually have to either increase the Strength to at least 0.79 (up to 0.99) or use the Enhance feature to buff up an existing image you've generated (which can include a separate Strength slider as part of what determines the degree of enhancement, should you wish to try and finesse it.) Later images in this batch show the results of playing with the Strength slider as well as the Enhance feature.
These images also show that if you provide an image with no background as a base, you have to specify the kind of background you want and pray (You'll usually have to increase the Strength as well to get satisfactory results, as otherwise the AI won't be able to deviate enough from the plain white background that it already sees.) You can do this via the tags you incorporate. For instance, I used tags such as field, blue sky, and later sunset, camp, and tent to get the AI to generate some basic backgrounds for Mallory to romp around in. The more images I generate of a character, the more backgrounds become a necessity for me, as you can really only do so many interesting things with a plain white background and a waifu. Even adding some basic colours and patterns as tags (specifying 'background' or 'theme' with each one) can flesh out the resulting images just enough to make them feel whole and not underwhelming. Sometimes the sparser backgrounds NovelAI generates can be fun, too! Take its tendency to have small background 'squares' as part of the images featuring Mallory, which gives those images a bit of stylistic zest while also making the white space a little more meaningful by association.
Finally, the Mallory session is just the first of many examples of how NovelAI has its idiosyncrasies regarding specific tags and won't always give you the results you want with them!
Example 1: I have found consistently that with the 'cat girl' and 'nekomusume' tags (the best ways I know of to tell Novel AI 'hey, that a neko') your output will often be missing ears and/or a tail, or on the flipside, it will have too many ears and/or tails! Some of these results can be influenced by the Strength slider (keeping the strength closer to the default makes it more likely to get the number of ears and tails right, but it's still not a guarantee,) but ultimately you just have to be prepared to work with these results when generating catgirl pics. Fortunately, there are ways of circumventing such issues, sometimes. Missing tails are the easiest to justify ('they're just hiding behind the neko, boiiii'), and if the catgirl's multiple tails are intertwined closely enough then sometimes using the Enhance and Variation features can blend them into one, though it might take a lot of fiddling to accomplish this. Ears are the more annoying factor: if your catgirl is missing ears in a result, it may be best to just try hitting Generate again, and also tweak your tags to more clearly specify that she should have cat ears (I suggest looking up ways to increase the weight of your tags if you need help with this.) On the other hand, if she has too many ears, you may be able to whittle them down to two by playing with Enhance and Variation, much like with the tail issue. Generally, I have found that it's easier to get rid of excess ears and tails than to try and get the AI to add them to fresh output, but there may be other ways around this dilemma that I haven't thought of.
Example 2: Fangs. Oh god, fangs. Be aware of what you're getting into when you include this tag is all I'm saying. This really doesn't come down to a specific trick involving tags or anything so much as it mandates you go into an image editing program afterward and clean up your characters' mouths. Sometimes, fangs will look natural. Sometimes, they'll look a little janky. And sometimes, it will look as though your character has an entire head of cauliflower stuck between their teeth. Enhancing images in very small increments can improve your results in this regard, but in doing so you also run the risk of making the output too noisy and/or blurry from repeated enhancing (something you'll notice in many of the images I end up saving because I end up giving up on them and/or backtracking to previous, less distorted results at that point.) If an image you've generated looks good overall, I find it easier just to save the image as is, then fix the fangs later. Same goes for other common little visual errors such as uneven pupils, lack of lines separating clothes from skin, and sometimes even too many or too few fingers (fixing this latter error can be time consuming, but it's worth it if the image looks solid otherwise!)
Example 3: Tan lines. Based on what I've read of other peoples' experiences with NovelAI, this is a notoriously unreliable tag. For starters, don't use it to indicate that a character just has a tan with no pale skin showing whatsoever. That's not what it's for. Use it instead to indicate that a character (for instance, Mallory) has clear tan lines where her skin gets lighter to indicate what areas of her are often more covered-up. But even in this context, be prepared to get a lot of weird looking results from using this tag. You can see many such results in the full folder attached, where the rest of the image looks fine, but the tag lines are just plain nonsensical, like, just impossible to justify, being the inverse of what they realistically should be on a body with tan lines, sometimes also melding with the lighting and shading for the hell of it and just cripes this tag is a frustration factory. I would go so far as to say this was the number one hurdle I had to overcome to generate good images of Mallory: her frigging tan lines! I used a combination of the techniques I mentioned with the previous examples to get the best results out of this tag: enhancing and generating variations rigorously, and then going into image editing programs to clean up wacky tan lines on images that were worth putting in that effort for.
Natasha
I had slightly more experience when generating these images, but I nonetheless stuck close to the default settings for the first few generations. I like doing this with a character I've never tried before, mostly to gauge initially how NovelAI grasps the basics of that character's design. In Natasha's case, the output was not only right on point but also quite stylistically charming at 0.7 Strength that I kept it at that for a while. Additionally, sometimes in limiting the Strength slider, you can keep the AI constrained by the base image just enough for it to successfully process parts of that image that tags won't necessarily cover. In Natasha's case, you can see this most clearly with her poofy tail and stocky figure, both of which were much less consistently present at higher Strength levels. Sadly, I couldn't figure out how to get the AI to represent her tail poof at all consistently past a Strength setting of 0.7, hence why almost all of the more detailed output of her is lacking that feature. While this discrepancy between her design and NovelAI's output was disappointing, I was willing to accept it given how well the utility captured the overall feel of her (aside from her whiskers and brown leather helmet, which were more understandably difficult features for it to grasp).
This session ended up being quick and fairly painless, with a good baseline of quality compared to other sessions from around the same time. I think the session was so consistent in its output quality because Natasha as a character has a bit of 'ordinariness' about her despite her in-canon celebrity status; She dresses all in beige and brown, and none of her clothes aren't too out there, which is a big thing when it comes to how accurately NovelAI will capture your character. Generally speaking, be prepared to see your character in slightly different outfits than you might be used to, as NovelAI can only do so much to capture all the nuances of an outfit you've designed. Clothes might end up being different colours than the ones you specify, and sometimes they might lack specific details even when you include them as tags (I've noticed that NovelAI has some difficulty with button-up shirts in particular, as in it just doesn't like them very much, but that's another story). This definitely happened a bit with Natasha, but overall it got the basic idea: cat girl with a long sleeve shirt, scarf, leggings, and boots. Most of the time these articles of clothing varied between shades of brown, orange and green, which I felt suited Tasha just fine. The helmet, as mentioned, was a bit less consistent, but again, I was willing to accept the variations in its design given that expecting NovelAI to consistently represent the same specific helmet design would have been expecting a bit much, and the overall feel of Natasha's design was still there.
The canyon background and colours were inspired by this piece: https://www.deviantart.com/axlwisp/art/Origin-Project-Natasha-612078341 though the image itself was never used as a base (I just figured NovelAI might possibly have some trouble representing a sandstone pterodactyl being decapitated by the weight of someone's belly.) Once again, the result felt very 'Natasha' to me, so not much tweaking was needed.
Leslie
Despite Leslie making up the majority of the final 'Best of' gallery, this session was the quickest and most efficient of the three. It was also the most recent, and so my comparative experience relative to previous sessions might help explain what I feel is a greater overall quality here. It likely doesn't hurt either that Leslie is probably the simplest and closest to anime waifu material as well. Very long hair, straight hair, white hair, maid outfit, and large breasts were some of tags I used to describe Leslie to NovelAI, and these tags also happened to be some of the tags with the strongest representation on booru sites, meaning the AI has much more data to draw from when interpreting them. Ultimately, this meant that the core elements of Leslie's design were always going to be pretty consistently represented in the AI's output, and so I was able to move on to higher Strength generations almost immediately without sacrificing quality. In fact, there was little reason not to max out the Strength slider here, as I was fully confident that NovelAI had seen enough big tiddy maids to get exactly what Leslie was about (please don't kill me, Leslie and/or Vespa.)
That said, the aforementioned inconsistencies with cat girl ears and tails were still a roadblock during this session. One additional inconsistency in this area that I haven't mentioned yet is the colour of these features, which, like with clothes, can be pretty difficult to hit consistently with your output. Leslie ended up having black ears and/or tail fairly often due to the AI getting a little confused by all the 'black' and 'white' descriptors attached to the tags I used this time, but I'm not sure there's much I could have done to prevent that, and the changes were so minor overall that I decided some such results were worth keeping if the image was of a high quality otherwise.
In a similar vein, observe also that fabled phenomenon of the 'floating cat tail': something NovelAI LOVES TO DEMONSTRATE and which you should thus be prepared to experience when generating cat girl images. I believe based on what I have read elsewhere that this tendency is rooted in people telling the AI again and again to generate very similarly shaped floating objects of a much less PG nature, but I can't say this for sure. Regardless, sometimes your cat tails just will not be attached to a cat girl (or even a cat, which also sometimes appear in cat girl images for obvious reasons.) Instead, they will float menacingly in the air next to the character, often just barely missing being attached to their butts, clearly just to spite you, the cat girl fan, for engaging in acts of pseudo-bestiality, you monster.
Also, sometimes the cat tail will be attached to something just out of frame instead of to the cat girl. Sometimes, there will be a second tail that does this, while the first is either properly attached, floating in midair, or doing the exact same thing. It's a toss-up! I'm not really sure how to mitigate these sorts of errors, and I've tried a lot of different things! Putting the tags 'multiple tails', 'two tails', 'second tail', 'floating tail', and 'floating p 3 n i s' in the handy Undesired Content box doesn't seem to help, based on my experience, but it doesn't seem to hurt either, so who am I to say don't bother with that approach?
Overall, this was a fairly simple session that yielded good results, some of the best I've seen even. Needless to say though, to achieve those results, I had to give up trying to create an illusion of modesty for poor Leslie and instead embraced the power of her absurdly unrealistic oppai for the sake of art- I mean science- I mean fanservice- shit.
