Mr Wisp
patreon
The Rest: February 2023
🕑 Added 2023-02-01 05:00:59 +0000 UTCHello! Before you proceed, please take the time to note the folder of 300+ AI-generated images attached to the end of this post! These images come from five separate sessions, each of which I'll describe individually, in chronological order. I'll also be including the images I used as the basis for each session to give you all a better sense of what NovelAI had to work with. For some basic notes on my AI image generation process as well as my intent/goals with the Re-AImagined content, check out my previous post, The Rest: January 2023, where I go into those things in detail.
Arciella Session #3 - Early November 2022
I refer to this as Session #3 because it represents the third session of image generation I undertook involving Arciella specifically. Coincidentally, this is also the third session of image generation I ever undertook. Arci was the first character I tried replicating using NovelAI, meaning the works you'll see involving her in this and in future months' galleries are the oldest out of all the Re-AImagined content. I'm not sure this is immediately apparent from this set of images because NovelAI did an excellent job capturing Arci's character, save for understandable omissions such as the spore caps on her belly. The Arci sessions were generally very smooth and a pleasure to work through, which is undoubtedly part of the reason I decided to devote more time to NovelAI content afterward.
Note that the lack of spore caps on Arci's belly in any of the AI-generated images of her does force me to be strategic about how I present these images. For future sessions, I would like to add these spore caps by hand to the top quality images since they are a fairly prominent part of Arci's design. Doing so will take some time and finesse depending on the detail of the base image in question, so for this session I opted instead to limit top quality selections to images where Arci's belly is covered, making it less apparent that her spore caps are missing.
Choices like this will factor into future content generated based on other characters of mine who also have less common features that NovelAI won't know how to represent. In some cases, the features are either complex or insignificant enough that adding them in by hand isn't worth the time: this was the decision I came to with Natasha's whiskers and extra-fluffy tail in last month's gallery, for instance. In others, the features are so integral to the character's design that I may not be able to generate a baseline of reasonable quality images for that character, meaning that for the time being I won't be able to feature them in these galleries. This is currently the case with two characters I've tried so far: Kira and Linda. Kira's design is too bizarre and unconventional for even NovelAI's furry model set, whereas with Linda the quality of the images is hampered by NovelAI having a tough time generating good looking lamia tails. Playing with the tags and settings might yet make it feasible to get some good image sets of Linda, tail and all, but at least in her case if I don't make much headway I can always just start with her human form and go from there.
There are a couple of idiosyncrasies with NovelAI that I feel are worth pointing out with this Arci session.
First is the common misalignment of wall, floor, and furniture lines. You can see evidence of this in many of the images from this session: sometimes the background gets disrupted by the character being in it and forgets what it's doing by the time it comes out the other end (that's my somewhat cheeky interpretation at least). The result is that something like the parts of a chair or table will be misplaced and won't flow cleanly into one-another. It can be difficult to fix this issue while preserving the qualities you like in the rest of the image, but I find that generating variations is one way to fish for a more 'structurally sound' version. Generating more than one image in the first place is another way to circumvent this issue, as it gives you more immediate choice, easing the pressure of trying to fix a wacky image. From my experience in general, it's a lot more enjoyable to generate multiple images more often instead of trying to get blood from a stone: i.e. spend hours trying to enhance one image. Even if you really like certain aspects of the image, it might not be worth the effort to try to fix everything about it. Granted, sometimes you can end up generating something unique and creative by putting in the work to tweak an imperfect image, so I guess just listen to your muse if you feel up for the challenge.
Second is that NovelAI has some unique culinary sensibilities. In other words, sometimes the food it generates can incorporate aspects of other tags, which can result in some unique looking dishes. I experienced this phenomenon with these generations, as many dishes ended up having candles and flames in them, which was very pretty if not a little bit concerning. Little mutations such as this can usually be fixed by enhancing the image with a moderate strength setting, or even just a little depending on the smallness of the detail. Sometimes you can see NovelAI tending toward such mutations before it gets there through the way details shift and transform when enhanced. Enhancing the same image multiple times is the best way to see this in action, and it can lead to some unique results if you let it follow its own logic. I ended up doing just that with some of the generations from this session, because ultimately the cool factor can be worth the suspension of disbelief. Besides, Arci's observant enough that she wouldn't actually eat a lit candle just because it was on a plate with a bunch of food, right...?
Anya Session #1 - Mid November 2022
Please note that this session refers specifically to the 'Anya, the Rotund Model' (Set 2/2) post. See the attached file 'Anya Session 1 Base' for the image I gave NovelAI for these generations.
You'll notice a drastic difference in look between this and other sessions. Most of these images were generated using a strength setting between 0.7 and 0.9, which has its pros and cons. To start with the obvious cons, these images don't appear as smooth or refined as others. They feel much nearer to my own style, but with some recurrent issues. But to be fair, these issues were due just as much to my own inexperience with NovelAI as they were to the limitations of the utility at this strength setting. For one, I overenhanced a lot of these images to try and make them better, which resulted in things like the backgrounds being oversaturated and the overall piece being 'blurrier' (I've tried to avoid reaching this point with image generations since this session.)
Another major thing I learned from this session is that the area between 0.7 and 0.9 on the strength slider can be a bit of a crapshoot. 0.7 is usually set as the default for image to image generations for a reason: it might not produce something completely fleshed out and awe-inspiring, but it will give you a sense of how NovelAI understands the base image while still generating something different enough to be interesting and sometimes worth keeping. There usually won't be a background unless your base image has one, in which case the background might be there but won't be very fleshed out. I tend to use this setting for more basic, cartoony, or abstract generations, and in this case I believe my intent was to try and preserve some of the style of the base image- an approach I've moved away from since then as it generally results in lower quality image sets. On the other side of the scale, strength settings of 0.9 and up tend to be where you want to end up. NovelAI can draw more from its existent knowledge base at these settings, meaning the baseline of quality is higher, at the expense of the output being more generic. I've had some success with image generations from everywhere along this spectrum, but less consistently than I have from just setting the thing to 0.99 and staying there.
Simply put, 0.7 can be a good, simple starting point if you want NovelAI to stick to the style of the base image but still produce something meaningfully different, and 0.99 seems to have the strength to generate something with a style and construction that's very different from that of the base image you provide, but with a good standard of quality, whereas 0.78 or 0.84 for instance might generate an uncanny in-between that tries to do too much and doesn't often succeed. This is where limbs and other body parts can get warped and even disappear entirely. Backgrounds can be detailed but sometimes aren't refined enough and just end up being ugly. Same with the characters. Like with basically every setting in NovelAI, there's potential to generate some unique and exciting stuff in this range, but more often than not I find it to be more of a headache than it's worth.
That said, if you don't experiment constantly, you don't know what you're missing, and I have certainly found that certain base images and tag combinations do better at certain strength settings. This is a potential pro of not going straight to 0.99 and instead feeling out how NovelAI reacts to the base image at lower strengths. Stylistically speaking, you can get some much more unique output this way, whereas keeping the strength at 0.99 all the time will result in overall more generic looking stuff.
So what does this all mean for this session? Well, that comes down to the subject matter, or more specifically, the fetish content. The terms I use to describe my art vary between commonly used and less commonly used tags, except in very niche communities that focus on these less commonly used tags. NovelAI draws from booru sites mostly, and understands some tags very well because it has a large selection of data to draw form with them. For example, the tags 'plump' and 'belly' tend to be the most effective ways to communicate that a character is fat or has a chubby tummy. These tags work well even at the highest strength settings, and so they are my go-tos when I just want the image generations to look good. More niche tags include 'big belly', 'fat', 'obese', 'feeding' and 'fat rolls'. These tags can produce the results you want, but it will take some work, and janky looking generations are common at all strength settings. For example, NovelAI has some weird tendencies with the tag 'big belly' where, if you try to emphasize it too much (by using curly brackets as so: {{{{{big belly}}}}}, ) or overenhance the image, the output will start looking very same-y. Specifically, the belly will usually turn white or grey and lack a navel, and a thick vertical line will appear down the center of it. The belly may even become detached from the rest of the body, appearing instead as a large beanbag hovering in front of the character. Furthermore, having multiple similar tags can cause them to clash, which is certainly true of fetishy tags. I've had image sets be plagued by multiple belly buttons, oddly divided bellies, elongated torsos with bellies stacked on top of eachother, bellies growing out of characters' kneecaps, and so on. This kind of weirdness might happen with most tags to some extent, but it seems more likely to surface when the tag or tags don't have as much data associated with them, so be attentive to these trends when working with such tags and adjust their emphasis accordingly. Try to keep note too of what tags don't play nice with eachother, and what if anything can be done to remedy that.
And then there are tags that don't show up among the suggestions when you type them in, which means there isn't a significant set of data representing them. This doesn't mean, however, that NovelAI can't understand these tags, but that the chances of it translating them into a quality image generation vary from unlikely to nil. Sometimes it's worth it to experiment if there's nothing else that conveys what these tags convey. Examples include 'stuffed belly', 'vore', 'huge belly', and 'burp' (sadly, for whatever reason, NovelAI seems to think that 'burp' specifically means 'beginning to barf', which makes this a risky and awkward tag to try and make good use of, at least for my purposes.) From my experience, using these tags at the highest strength setting just drags down the quality of the resultant generations instead of representing the tags as well as I hope. This is a reason why setting the strength between 0.7 and 0.9 can be useful. Because the styles of these generations are usually less generic and more experimental, less-understood tags tend to appear less jarring and might actually blend in decently with the others.
I used the tags 'stuffed belly' and 'burp' during this session to see if I could communicate to NovelAI that Anya's midsection should be round and tight rather than fat, but not necessarily pregnant- and also that she be burpin. It seems to me that NovelAI did manage to take into account these more obscure tags, with varying results. While the 'burp' tag is still minimal in its impact (unless you want to see some barfing, which I personally do not), I get the sense that there might be some potential to do some more belly stuffing themed generations once I learn more about what sorts of other tags might work best to complement this theme. The overall quality of these images isn't great, as the content didn't translate properly at higher strength settings and so I kept the strength relatively low, but I do think these are still among the most successful stuffing-themed generations I've been able to produce so far. Anya actually looks stuffed rather than just fat in many of these generations, and the belly shapes are fantastic, which does to some degree make up for the cruder style as a whole. Expect to see further attempts at better stuffing content in the future.
Anya Session #2 - Mid November 2022
Please note that this session refers specifically to the 'Anya, the Revered Model' (Set 1/2) post. See the attached file 'Anya Session 2 Base' for the image I gave NovelAI for these generations.
I'll be briefer with this session since it's a lot more straightforward than the previous one. Here, the tags I used were much more up NovelAI's alley and so I was able to get more consistent results at higher strength settings- though I did sometimes keep the strength lower out of curiosity and desire to experiment. The tag 'reclining' got across exactly what Anya was doing, though I did fuss a bit over how to have her be propped up on one arm. Sadly, I wasn't fully in the habit of saving the tags I used for each specific session at this time, so I can't say what ended up doing the trick in the end.
I believe I used the tag 'painting' here to try and create the sense that this could be what a portrait of Anya done by a real painter might look like in Axu2 canon. I may end up sneaking one or two of these images into the game for that purpose, but ultimately I don't think any of them ended up looking too distinctly like actual paintings. Still, generating this kind of output after the rockier first session was encouraging, and gave me a better sense of how the sum of the tags used to describe an image could affect the style and quality of the results.
As always, there was some typical catgirl jank to work through here. The 'detached tail phenomenon' that I mentioned in last month's notes was at play here, as was the disappearance of ears. Sometimes the tail melded with Anya's hair. Sometimes the colour of the ears fluctuated. Some of these things I just had to accept as part of NovelAI's 'artistic license', so to speak. Additionally, this was the first time I had tried landscape dimensions for image generation, which apparently are less reliable than those of a portrait. I'm not sure if this is why Anya's legs sometimes ended up misshapen or missing segments, but overall the quality was still good enough to make the dimensions worthwhile. I have yet to experiment more with landscape dimensions, but hopefully I'll be able to again at some point soon!
One more small tidbit worth mentioning: throughout all of these generations and those of the previous session I had struggled to properly convey the shape and colour of Anya's hair to NovelAI. In some regards this was never going to be feasible (I'm looking specifically at the way her hair parts in little 'rolls' or humps.) But thanks to some research I did, as well as a suggestion from my girlfriend who knows a lot more about long hair than I do, I came up with a couple of strategies for capturing the feel of Anya's hair with reasonable accuracy. Colourwise, my issue was that in terms of the suggested tags, NovelAI does not understand that 'blonde hair' can be a darker or more orangey shade of blonde rather than just yellow (or platinum blonde specifically, for whatever reason. Apparently NovelAI just thinks those are the only two shades of blonde out there.) This problem was solved part of the way through session #1 by putting the tags 'blonde hair' and 'red hair' next to each other and then deemphasizing the 'red hair' tag using square brackets. This is the inverse of the curly bracket trick I mentioned above: instead of making NovelAI emphasize the bracketed tag more, it subordinates the tag compared to those around it. The two tags ended up looking like this: blonde hair, [[red hair]]. This made Anya's hair just a little redder (sometimes a bit too red, though I felt that was more fitting than the yellowy blonde she'd been stuck with prior) and ultimately matched her actual hair colour a decent number of times, which was exciting!
The other hair problem I managed to solve was with Anya's bangs, which had been all over the place throughout session #1. By session #2 though, I managed to get them to behave a little more like Anya's actual bangs. Interestingly, I did this using a tag that NovelAI didn't recommend, but which my girlfriend did: 'side swept bangs'. The 'swept bangs' tag itself was one that NovelAI recommended and which I had been using, but it didn't convey properly the way Anya's bangs flow from one side of her forehead to the other. Adding the word 'side' to the front of the tag solved this issue and gave Anya just a bit more distinct Anya-ness for this and subsequent sessions. The moral of the story is twofold: 1) don't be afraid to experiment and tweak existing tags to see if they help NovelAI get a little closer to your vision, and 2) your girlfriend is always right. She'll probably agree with you on this if you ask her. Just saying.
Abby, Session #1 - Early January 2023
See the attached file 'Abby Session 1 Base' for the image I gave NovelAI for these generations.
This is one of the more recent image sets I've generated, and I haven't spent a ton of time on Abby compared to some of the other characters I've tried thusfar, which is why there are fewer images of her in this folder than of, say, A N Y A (who I still have a few more sessions worth of images of saved for future months). There are a couple of reasons for this.
Firstly, timing. I started the Abby generations at a time when I didn't have much spare money to spend on anlas, and I took a lengthy break from NovelAI after this session to focus on work and academics.
Secondly, well, it got her right on the first try! I didn't have a lot of nagging urges to tweak the tag selection and try different things to improve or change the result because the result was good immediately. Abby looked like Abby! This doesn't mean that I won't go back to her later, as I fully intend to do so, but rather it is to say that sometimes a character coming out perfect allows me to move on to another character rather quickly. Or, to look at it the other way, when a character comes out not quite perfect, and I think I can improve the results, I'll spend a lot more time generating images of that character, some of which will still end up being pretty decent images! This is one reason why I ended up with hundreds of Anya images as opposed to less than fifty Abbies.
The last reason is that I knew what I was doing by this point compared to when I started: not only did I have a better grasp of the intricacies of NovelAI and how to get more consistent output, but specifically I knew how to generate catgirl images after hours and hours of fiddling with Anya and Mallory... and Natasha, and Leslie, and you get the picture. Abby had the benefit of coming after all those characters, so I was able to spend a lot less time thinking and experimenting. Granted, this means the Abby content is pretty run of the mill and not too stylistically varied, but for the small amount of images I ended up keeping, the ratio of quality versus quantity is one of the highest I've had so far!
Some specifics from this session: I started at 0.7 strength, found that NovelAI was generating peak Abby from the get-go, but decided to stay at this strength setting for a little longer just because the simpler and more line-heavy style fit her nicely. This is a case where transitioning to 0.99 strength wasn't all upside, as it introduced some large, dynamic, distracting sets of challenges that forced me to adjust the tags and also my expectations regarding output. Yes, I am talking about boobs. With Abby, I decided to be somewhat lenient here since she isn't flat-chested, just smol. But (incoming hot take) the problem with generic waifu images made with NovelAI is that you really have to scream at the utility not to make the breasts gigantic, or else it will gladly assume that's ok. Look, I'm not going to act like I don't know why it's developed this behaviour, nor am I going to act like I'm not part of the problem (you all saw the Leslie pinups from last month). But obviously this can be inconvenient when you just want to generate some images of female characters with realistic chest proportions. I find even with characters who I consider busty relative to the average female that I still often end up navigating to the 'Undesired Content' box and inputting the tags 'Huge breasts' and even 'large breasts' just so NovelAI doesn't keep giving said characters the Jessica Rabbit treatment. In fact, I did just that at the start of the Renee sessions, as well as specifying 'medium breasts' in the actual tags, and the images you'll see from those sessions are the result of that. Granted, I do get the sense that adding the tag 'plump' doesn't help with this particular dilemma. Anyway, as far as the Abby sessions went, she ended up bustier than normal, especially with the later, high-strength generations. C'est la vie!
Now let's move away from the topic of pecs and talk about guns. Not pectoral guns. Gun guns. ...Ok, well, space guns in this specific case. The point I'd like to make here is that adding gun tags can be risky with NovelAI, unless you don't mind a very ambiguous looking firearm. This comes down to the little details the utility generates, which are very rarely all neatly and cleanly arranged for maximum realism. Think back to what I said in the section on Arci regarding furniture. The same is true for anything that has a rigid and precise design, which includes guns. Fortunately, this isn't as big of a deal when the guns you want your character to be holding are fictional guns without a particularly specific design. I used this to my advantage to generate firearms for Abby that could be a number of things: a crazy space gun maybe, or perhaps her signature hookshot? Maybe she likes to modify weapons (note: this is not really a 'maybe.' Abby does in fact like to do this.) Regardless, having the guns she's holding vary and be somewhat vague in appearance wasn't such a problem in this case, which certainly helped with the overall quality of the output in terms of what I wanted to see. That said, there will always be instances where the gun just looks wrong no matter how you're planning to justify it. Just like with furniture, that's pretty unavoidable, but at least there's a bit more of a margin for 'creative error' in this case.
One more slightly embarrassing note regarding these sessions: I tried repeatedly and with increasing emphasis to get NovelAI to respond to the tag 'armpit hair.' Why? Because Abby doesn't like shaving those. No, really. It's in-game lore. I'm not a weirdo.
...Ok, I am a weirdo, but I still did it for the sake of accuracy! I also didn't expect or want to see much hair, but I figured that wouldn't be an issue because NovelAI wasn't likely to get out of control with the tag, and in that sense I was correct. There are a couple of images in the attached folder where, if you squint, you can see the tiniest hint of pink fuzz in Abby's armpit. That is the final result of my attempting to get it to respond to this tag. To the shock and amazement of all of you I'm sure, NovelAI does not seem to want to add pit hair to waifus. But I do give it bonus points in this case for assuming it would be pink. Of course, an awkward side effect of me emphasizing this tag so much was that, while there wasn't much armpit hair in any of the images I generated, there ended up being a whole lot of armpit. This is why Abby's shoulders and armpits in some of the later images look a bit... like airport landing strips. Apparently gigantic, sprawling armpits are fine and dandy waifu tech. Just don't you dare imply that hair grows in any of them. Disgusting.
Renee Sessions #1&2 - Late January 2023
See the attached files 'Renee Session 1 Base' and 'Renee Session 2 Base' for the images I gave NovelAI for these generations. The first image was used for the initial generations featuring Renee in a stairwell. The second image was used for everything that came after.
These are the most recent sessions I've had with NovelAI to date, and I'm very happy with the quality overall. Things started out a bit more 'psychological thriller' with the stairwell setting, which wasn't entirely my intention. The base image wasn't much help to NovelAI here, and it turns out the tag 'spiral staircase' isn't the most consistent tag anyway! Granted, stairs just seem to be one of those things you should avoid trying to include in your image generations, as they very rarely look realistic or even physically possible. But if you want some cool, MC Escher style paradox-staircases then congratulations, you've found your image generation paradise. Just don't expect them to be as precisely and carefully paradoxical as Escher's. NovelAI doesn't do precise.
I don't have much else to say about the first session, aka the 'Staircase Session' other than it nailed Renee herself! Adding emotion-based tags isn't always a good idea with NovelAI, as it tends to translate those emotions weirdly into images (just try adding tags like 'happy' and 'sad' to your image generations and note the difference). But here I was able to use the tag 'sleepy', with some deemphasis using square brackets, to give Renee her light sleepy vibe. I also chose not to worry much about how her hair turned out with either of these sessions, as long as it was black, curly, messy, and not too long. Black is a pretty difficult colour for NovelAI to get wrong, and Renee ended up being a lot easier to represent than, say, Anya because of that. One issue I had early on was that the 'freckles' tag was a bit too prominent, but square bracketing that tag helped make this less of an issue.
As mentioned above, these sessions were a case where NovelAI decided it wanted the boobs to be gigantic, and the majority of the time I spent fiddling with tags was actually spent trying to communicate that Renee's chest was not that massive. But this does speak to the fact that proportions will always vary between image generations despite your best efforts, and so for the sake of your sanity it's easier to just accept this fact and focus on more accurately conveying other aspects of the character rather than get frustrated because the AI likes its waifus extra thique.
With Session #2, aka the 'Sweater Puppy Session' specifically, NovelAI also seemed to like to imagine that Renee was wearing multiple layers of sweaters, which I decided was fine. In fact, I might just make it canon because the more I think about it, the more it makes sense. Of course she wears multiple layers. She's a vampire. Most of the time, she's fricking cold. And sweaters are comfortable, so yeah, she'd rock a sweater cocoon in a heartbeat. This also kind of helps me accept the big booba dilemma too, since I can pretend that part of their disproportionate size is just due to there being more padding over them! It's the sweaters, clearly! Yes, yes, I like that. It's lore now.
As for why this happened, it may have just been because I included both the tags 'turtleneck sweater' and 'sweater lift', so NovelAI sometimes assumed that meant multiple sweaters...? That's my working theory anyhow.
Worth noting is that I added a red background to the second base image for this session because otherwise the backgrounds would have been primarily white, and I didn't want that. The setting this time was supposed to be Renee's dank, bloody basement, not Renee's conspicuous white void. Fortunately, the plain red background and a few descriptive tags such as 'underground', 'brick wall', and the ever-descriptive 'dark' were all NovelAI needed to construct this setting with reasonable success.
I found overall that Renee was one character who benefitted from a strength setting of 0.99 because her most prominent features aren't all that uncommon, especially not among waifus. Thick eyeliner? Check. Messy black hair and pale skin? Sure. Sweater puppies? Of course. Thick thighs in leggings? Yes, welcome to the internet. Catgirl? Expect the tail to sometimes multiply or detach and float in midair, but otherwise, you're all set. The only element of her character I feel was a real struggle was her coffee. I gave up on having her hold a coffee mug after Session #1 because NovelAI rarely got it right. I'm hoping to find some way to bring out Renee's love of the brew for future sessions, without sacrificing quality if possible. We'll see how that goes.
