Mr Wisp
patreon
The Rest: June 2023
🕑 Added 2023-06-01 04:00:04 +0000 UTCHello! Before you proceed, please take the time to note the TWO folders containing a combined 350+ AI-generated images attached to the end of this post! There are two folders this time because the average image size has increased significantly, which pushed the zip folder size past what Patreon could handle for an individual file. These images come from three 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 'The Rest' posts, where I go into those things in detail.
Before we get into the first session's notes, I wanted to briefly address the increased average image size. When generating images with NovelAI, the rule of thumb regarding size is as follows: the larger the image dimensions, the more anlas it will cost to generate, enhance, and do just about anything with the result. While I love larger images, for the sake of practicality, it doesn't make sense to use larger dimensions for each and every image generation, especially when there's no guarantee that the result won't be worth the investment. This is where it's relevant to bring up a facet of NovelAI that I haven't discussed much throughout these notes, that being its upscale function, which allows you to increase the size of a generated image while maintaining a reasonable image quality. This function allows you to start by generating smaller images and then enlarge only the best results. I avoided using this function for several months since I wanted to focus on learning the ins and outs of image generation, variation, and enhancement first, all of which required anlas. Since upscaling also requires anlas to use, I didn't feel it was worth constantly using, and if I wasn't constantly using it, I worried I'd end up with a distracting amount of variation in image sizes for each session. I began to rethink this habit around the time I was working on the Raziel Jungle session a few months back, partly because the dimensions of these image generations were more square and thus ended up smaller than normal, which made me feel they should be just a bit bigger. I ended up only upscaling the results I thought were the most promising, but I ended up really liking some of the smaller images that I hadn't upscaled as well, so the result was still that the image sizes for that portion of the gallery weren't consistent. Still, looking at it after the fact, I felt that the gallery benefitted from the larger image sizes enough for me to give the upscale feature more consideration in the future. Fast-forward to my more recent sessions, which include all the sessions featured in this month's gallery, and I'd decided to try and have all the main gallery images be upscaled. While this is more anla-intensive, I felt by this point that I was able to get away with it because I understood enough about NovelAI that I was generally able to produce better quality results quicker. This still meant starting small and only upscaling the best results, but in cases where I felt later on that a smaller image I hadn't upscaled was worth using, I tried to leave some anlas open so that I could go back and upscale that image. I didn't end up doing that with any of the images from these sessions, as I had a good enough eye for what deserved to be upscaled that I ended up agreeing with my own upscale decisions after the fact (though if you, the reader and loyal patron, think there are some images among those not featured in this gallery that weren't upscaled but maybe deserved to be, I'd love to hear your thoughts!) All in all, it ended up that having the entire gallery feature upscaled images wasn't as much trouble as I thought it would be, at least for this month's sessions, so I'll continue with this practice in future sessions, barring some unforeseen exceptional circumstances where upscaling the best results somehow becomes less feasible, though I doubt such circumstances will come up. TL;DR - Image dimensions be bigger from now on, enjoy!
Ashley Session #1 - Late April 2023
This session was done immediately before the Ashley session from last month's gallery, meaning it was my first time doing image generations featuring this character. I have a lot of varying thoughts about how this session went, but in terms of the overall image quality, I think it went pretty well. Designwise, Ashley is in some ways simple to describe (humanoid, young woman, skinny, pale skin, red eyes, pointed ears, somewhat bony figure.) so it didn't surprise me that NovelAI was able to produce good results without much trouble. That said, there are a couple of elements of Ashley that aren't as intuitive, such as her very thin neck and rounder head shape, and I struggled to convey these traits properly. This wasn't a huge deal since the results were still recognizably Ashley, but since these traits are a part of her anatomy rather than stylistic choices, it's still a little frustrating that I had to compromise on them. If/when I return to Ashley for future sessions, I'll be attempting to figure out how to better convey such cabra-specific physical features before anything else.
Another area I had to compromise on here was that of the setting. I did include tags like 'floating islands', 'cliffs', and 'above clouds' in an attempt to get NovelAI to generate backgrounds that resembled Axu2's Highlands, but I never felt that the background quite got there, though I can't complain about the results either, as they ended up being generic enough to do the job just fine.
Key tags this time around include 'bored' and 'eyeliner' (to help replicate Ashley's impossibly dark eyelids), '{{big twintails}}', '[big head]', '{long neck}' and 'big eyes'. These tags got the results close enough to Ashley's actual appearance that I felt like I was looking at Ashley rather than someone doing an Ashley cosplay, which is always a good sign that I can move on to tweaking other tags.
Since Ashley is just about the skinniest character of mine (not that there's a ton of competition in that department) I used a couple of bone-related tags to help emphasize her relatively unique figure: 'Collarbone' and 'hip bones', namely. These tags put in a lot of work in my opinion, as they show up to some extent in the majority of the results from this session and helped set Ashley apart from fellow skinny character Amber, for instance. Granted, Ashley's outfit complemented these tags perfectly, even if it ended up typically being a lot skimpier than Ashley would find practical. The following tags informed how NovelAI generated said outfit: 'ninja', 'vampire', 'bare shoulders', '{purple leotard}', 'skin tight', 'long sleeves', 'fishnet legwear', 'purple stockings', 'slim legs', 'purple cape' (note that the single quotation marks weren't included in the prompt and are only added here to distinguish the tags from everything else.)
To some extent, involving the tags 'leotard' and 'skintight' meant that the outfit looking skimpy was a foregone conclusion, though in some cases NovelAI took a fair bit of creative license in just how skimpy it could make it, resulting in the leotard in particular being awkwardly low cut in a lot of the results (the 'bare shoulders' and 'collarbone' tags may have also contributed to this.) If I revisit Ashley's ninja outfit for future sessions, I'd like to see about having it look more consistently practical and less blatantly skimpy, not because I think any of my patrons mind, but because it's pretty out of character for Ashley to wear something 'sexy', especially when it's the outfit she uses for stealth missions. This is definitely another case of NovelAI being very used to generating certain types of content, so to speak, but there are still probably ways to dial it back a little when it comes to characters who aren't about the sex appeal. It might also mean I need to do some more research on more practical outfits and what tags best convey them to NovelAI. As a fetish artist, realism obviously isn't my top priority, but matching a character's aesthetic to who they are as a person is pretty close to it.
This session had three distinct phases: the first was the 'DC Superhero' phase, which includes images from 'shl5' to 'shl8' in the second attached folder. It featured the tag 'dc comics' since I was initially interested in giving a darker and grittier visual style to the session. It ended up making some decent results, including one of the images featured in the main gallery, but shockingly, it also made Ashley look a bit too much like a DC Superhero in just about every way, which wasn't a vibe that quite fit her. I don't think it was a bad fit per se, and it may be that I revisit 'Batshley' far in the future, but I wanted to try some different ideas, and so the dc comics tag was dropped fairly soon into the session.
The second phase was the 'Ashley holding Katana' phase, which includes images from 'shl9' to 'shl37'. The trend throughout this phase of the session was me being very dissatisfied with NovelAI's ability to generate good looking weapons, which is nothing new. I've discussed these sorts of issues in previous months' notes, so I won't rehash them here, but suffice it to say that having Ashley hold a katana was often more of a hindrance than a boon to the average quality of these results. In some cases the jankiness was mild enough that I felt I could either smooth it out via editing the image by hand and/or play off as a stylistic quirk. Sometimes the katana looked more like a dagger, or a switchblade, and in those cases I had to accept that this was fine because hey, Ashley probably also has one or both of those weapons on her person at all times, so the lore holds up. Sometimes the katana had no hilt or sheathe; in fact, it was abnormal if there was a sheathe, but again, this is nothing new. I don't have much else to say about this phase of the session as, aside from weapon tag jankiness and creative outfit decisions, it went pretty smoothly. I did the usual progression from 0.7 to 0.99 strength to start out this phase, though I'm not sure I really needed to since it was clear from the beginning that NovelAI wasn't going to have any trouble figuring out how to Ashley. I do like a couple of the lower strength results, enough that I featured one of them in the main gallery, but the higher strength results were clearly the more interesting and fleshed-out of the two groups, so I was content to stay near 0.99 for most of the session.
The final phase of this session was the 'Ashley standing in a lake' phase, which started with Ashley holding a weapon and then transitioned into her holding a vial of blood. This was the stage of the session where I felt I had enough quality images of Ashley already that I could get away with adding belly tags, which can sometimes tank the quality of generated images if you aren't careful. In this case, I didn't emphasize these tags (or rather tag; {{big belly}} was the only one I used) much, and so the quality wasn't too adversely affected. My vision for this phase of the session was that, after a night of patrolling the Highlands, Ashley typically retires to a secluded lake, uncorks a vial of blood or three, and treats herself a bit. ...Look, she may not be a glutton on the level of some of my other characters, but she still has her vices. Fetishy leanings aside, I'm happiest with this phase of the session because the setting fit well with the overall mood, especially the colour palette. The lake ended up looking ethereal and really intriguing in a lot of the results, which helped emphasize the subtle fact that Ashley isn't human, even if she's pretty close. I don't mean to position her as a mysterious, exoticized 'other', mind you, just a person from a race and culture that have developed in their own unique way based on their environment- in this case, being hemovores (beings for whom drinking blood is a necessity) and somewhat nocturnal, while also living in a remote geographic area with a very insulated and tight-knit society. My intention, as with all characters like her, is to depict Ashley as someone whose daily life is shaped by all these culturally-specific realities, and whose actions and motivations are just as valid as anyone else's. I'd like to develop this general vibe more and carry it over to Axu2 where possible, as giving Westhaven and its people more of their own fleshed-out culture is something I'm always interested in.
Claire Session #1 - Late April 2023
Here's one of a few characters I didn't expect to be including in one of these galleries any time soon! This was definitely a more experimental session than anything else, and the results aren't perfect, but they are a whole lot better than what I expected! I wasn't planning on using any of the images I generated during this session when I began it, with my assumption being that none of them would look remotely like Claire. There was some precedent for this assumption: I've had a few sessions at this point where the results were all more or less unusable, and unsurprisingly, all of them involved older characters of mine whose designs aren't particularly conventional or intuitive: Kira is the best example of this, as well as Axi to a lesser extent (though that session got pretty close to producing good results so I haven't given up on her just yet). In Claire's case, I figured the red skin, vaguely batlike ears, and flipper feet would be enough to throw NovelAI off, but there were a few things I didn't consider when I predicted this: for one thing, I'd done plenty of sessions involving characters with nonhuman skintones before (though most of those sessions have yet to be featured), and aside from a few glitches, NovelAI was able to get their skintones right more often than not; for another thing, Claire's ears aren't all that different from typical anime elf ears, or at least they're close enough to me that I didn't mind compromising on the specifics of their shape and size as much as I thought I would; and finally, whether or not Claire is wearing her stupid-ass flippers is a moot point if her feet are always just out of frame (:
With some apologies to Claire's flippers, I am serious. Not having to worry about her feet, which are easily the least intuitive facet of her design, made the rest of her design look completely achievable by comparison. Yes, I had to compromise on her ears a bit, but I've seen bigger design changes be attributed to stylistic difference, so I was fine with it. Once I got over these hurdles, the only other part of her design that NovelAI really struggled with was her eyes, specifically her lack of proper sclera. This was something I ended up having to fix by hand, but that wasn't a huge ask since I typically have to spend a fair bit of time cleaning up the eyes of every character anyway.
Now, I realize I might be giving off the impression that I solved the entire puzzle of how to generate quality Claires in NovelAI, but that's not completely true. There's still a lot of work to be done before I'll be able to include her in one of these galleries as more than a small bonus section, because the majority of the images I generated during this session were just too much work to fix up. Part of the problem was finding the right settings, as NovelAI didn't handle Claire like it does my more intuitively designed characters: at 0.7 strength it could produce her exact skintone, hair colour, and pose, and it could even sometimes give her glasses that didn't look completely atrocious, but the results were pretty bare bones and not enough to base a section of a month's main gallery off of. On the other hand, at 0.99 strength, it would get her skintone right most of the time, but it usually wouldn't get her hair at all, either making it a different shade of blue or a completely different hairstyle (I was a little more lenient with the latter problem since characters can grow out their hair and such, but it still wasn't ideal), and the glasses would be missing entirely. In fact, the glasses tended to disappear around 0.90, which ended up being the best midway point between image quality and accurate representation of her design. That's why most of the images in this gallery don't include Claire's glasses, which isn't ideal but ultimately not a dealbreaker either. She could wear contacts if she wanted to. She'd just hate every minute of it.
This wasn't a particularly long session either, as I ran out of anlas midway through, and I wasn't entirely sure where to go with it anyway. Looking back on it now though, I definitely have a few ideas, and I'm hoping to make some more progress in a session in the near future. But as far as this month's gallery is concerned, it meant there wasn't much to choose from, hence why I've included only a handful of Claire pics this time. Still, as I've said, even that small amount is a lot more than I'd expected, so I consider it a win. I'm going to gloss over the fact that, as usual, NovelAI struggled with a character's weapon more than it did with the character itself, though it does make me wonder if there's some way to specify with tags what the weapon should be doing in a given image. Even some negative tags like 'bad weapon' or 'impossible weapon' might help a little. Imma look into this next time it comes up.
One final note on this session, and perhaps the most apparent difference between Claire's actual design and NovelAI's take on it, is that her skin ended up a lot 'busier' than it actually is: by that I mean NovelAI added a lot of stripes, splotches and other patterns to it. If I had to hazard a guess as to why, I'd say it has something to do with the 'salamander girl' tag I used to try and help convey what Claire's skin was like. This wasn't a stretch on my part since Claire is canonically an amphibian of some description, and I've used the term salamander when tagging images of her before. What I didn't expect was for NovelAI to take that idea and run with it by giving her a bunch of salamander-esque skin patterns. If this were any other character, this development would have been a problem, but because even I have no idea what Claire is supposed to be, I'm just grateful for the effort on NovelAI's part. What's more is, I actually like the design choice it made by including these skin patterns, and I'm seriously tempted to revamp Claire's design a little bit to include them. She already has freckles, and her design as it stands isn't particularly busy, so some more amphibianesque skin-patterns would probably flatter her, if these images are anything to go by. I'm not saying this design change will happen in the immediate future, but it is something I'm genuinely considering, and from a thematic standpoint, the fact that it was essentially 'suggested' to me by an ai is... weirdly appropriate for Claire.
Nell Session #1 - Mid May 2023
This is a very recent session that I did as soon as I saw that the vote for this month's gallery was going to be near-unanimous. I've featured a couple of Axu2 party members in these galleries already (Abby, Arciella, Evelyn, and Ashley), but among those I haven't featured yet, Nell is somewhat of a fan favourite, so when I saw how skewed the vote was in her general direction, I figured it was time to give the people what they wanted. I don't have a ton to say about this session overall because it was just about the simplest one I've done so far. I knew ahead of time that it wouldn't be difficult to make Nell work in NovelAI, and it wasn't. This isn't to say that the results were immediately perfect, though if you note the average quality of the images in the accompanying folders then you might agree they were pretty darn close to it. If anything, the quality went downhill rather than uphill after a while, and that's because I by that point I had started playing around with tags I knew were a bit dodgy ('obese', mainly. Avoid that tag specifically.) The fact that I felt comfortable doing so is evidence of how well the session went overall, as despite the relatively low number of images generated, I quickly felt like I had enough quality results to fill a generous portion of this month's gallery. Needless to say, I'm looking forward to pushing the boundaries of what I can do with certain tags in future Nell sessions, now that I know the baseline of quality is so high.
There are a few caveats here. One is that I didn't spend much of the session at 0.99 strength, but rather somewhere around 0.91. This was in hope that reducing the strength somewhat would keep NovelAI focused on the base image, specifically with regard to the size of Nell's ponytail. This sort of worked, but it was a tall order given the ridiculous size of said hair. I accepted pretty swiftly that NovelAI was never going to be able to make it big enough (see 'nll3' for an example of Nell's hair being so massive that NovelAI assumed it was actually the sun or something), but it did a pretty darn good job nonetheless. '{{{big ponytail}}}', '{{{wide ponytail}}}', 'giant ponytail', '{{long ponytail}}', and '{big hair}', were the tags I used to try and coerce the ai into making Nell's ponytail as comically large as possible, and hey, sometimes it really did get there! See 'nll12' through to 'nll15' for a series of examples. Ultimately, I had to accept with this session that Nell's ponytail was going to vary drastically in size (I had experienced this already with Ashley's similarly large twintails, so it wasn't a huge surprise), and I doubt that is something I'll be able to fix in future sessions. I haven't entirely given up on it since there is clear evidence that NovelAI is able to process hair that big, but there's no chance of it ever being able to do so consistently, so I'm willing to accept a medium or medium-large ponytail if the rest of Nell is accurate.
Another major caveat is that Nell's eyebrows, eye piercings, and earrings as you see them in the main gallery were almost entirely added by hand. I included the tags 'long eyebrows', 'antenna eyebrows', and 'eye piercing' throughout this session, and while the first two tags may have helped push Nell's eyebrows in the right direction, the third had no impact whatsoever. Granted, I didn't bother emphasizing these tags since I figured the odds of them working out like I wanted were low, and drawing them on after the fact was no trouble at all, so I decided quickly that it'd be simpler to do just that. Another feature of Nell that didn't really translate here (partly because I've never conveyed it all that adequately) is her spinal deformity, which causes her to, for lack of a better description, lean forward with her butt out like she's one eighth of the way into doing a squat. Physical deformity isn't something to make fun of, so I apologize for the comparison, but either way, this general pose is what I had in mind when I drew the Nell reference that I ended up using as the base for this session. I don't blame NovelAI at all for not picking up on this aspect of Nell because not only is it not very obvious just from looking at her (regardless of my imperfect understanding of how spines do and don't typically work), but I also wasn't sure how to go about emphasizing it via tags. I ended up relying on the tag 'leaning forward' and hoping it would do the trick, which I quickly realized was not going to happen. This is why the first few images from this session have Nell doing a somewhat awkward 'look at my tits' pose, which is fine as fanservice but not so much as an accurate representation of Nell. I eventually gave up on trying to get Nell's odd spine to work in NovelAI, but it's something I'm going to give another shot at next time, as even if it isn't absolutely essential to her character, it's not something I want to pretend isn't there either.
Here's another one: for most of this session I used some variation of the tag 'holding axe' or 'holding battleaxe'. I knew full well beforehand that this was going to result in some jank, but the inconvenience was still worth getting a few good results featuring Nell holding her signature weapon, or at least something close enough to it. That said, the next step is getting her to dual wield axes like she does in Axu2, which is going to be a next-level headache in and of itself. As you can see from the images in the attached folders, Nell ended up holding a lot of things that definitely weren't axes over the course of this session, none of which were intended, but some of which looked cool or on-brand enough to roll with. Special mention goes to the staff-like weapons she's holding conkeldurr-style in some of the variations of 'nll8', which I was tempted to include in the final gallery before deciding they looked a bit too much like either giant ski poles or walking sticks. ...Now that I think about it though, Nell not only deserves but could probably flat-out rock a pair of walking sticks that double as murder weapons. Gonna have to keep that idea in mind...
Alright, here's the obligatory 'she vore things so I added big belly tags' section. It's exactly what it looks like, folks. I used the tags {{{plump}}}, {{{big belly}}}, and 'pregnant' specifically to push the size and shape of Nell's midriff in this direction, and I think the results are about as good as I could hope to get from an ai that hasn't been trained primarily using fetish art. It's worth mentioning here, for those who aren't aware, that StableDiffusion is another ai image generation utility that has a few 'models' trained under these constraints (you can find them if you do enough research and frequent the right Discords), and the temptation to rely on these models for more explicitly fetishy results is real, but I tend to avoid doing so for two reasons. 1) even though it is free (which is a big incentive to use it over NovelAI) you have to go through more hoops to make the utility work, and even then it is buggier and less intuitive in several ways, and so more tenacity and general knowledge of how it works is needed to get the best results; and 2) because these models were trained using fetish art, this means that the results you're getting are much stylistically closer to the output of fetish artists, many of whom I consider inspirations, contemporaries, friends, etc. NovelAI's much wider image pool and lack of fetish-specificity means I never have to worry about my results aping another artist's style, but with StableDiffusion, it's a real issue. As an artist who can always just draw my own darn bellies, I feel like using models trained on art from the same circles that I swim in is a lousy thing to do, so I avoid doing it. I think it's a different matter when you're a fan who just wants to make some images in the style of your favourite artist, but there are probably many artists who would disagree strongly with that sentiment too, one way or another. Either way, I digress. I do have a few images generated using StableDiffusion, but their fetish content isn't very specifically tagged, and stylistically they don't seem to resemble the work of any specific fetish artist that I know of. Whether or not these images see the light of day will depend on how much research I'm willing to do just to make sure they aren't actually a little too stylistically indebted to anyone in particular. You can call me a hypocrite for worrying about this but not taking any issue with using tags like 'dc comics', and you'd probably have a point, but since fetish art communities are smaller, much more independent, and reliant on honour and integrity among artists in the absence of any written agreements or corporate backing, I think taking extra precautions not to step on any toes within these specific communities is very warranted.
With that seriousness aside, I'm overwhelmingly happy with the results of this session and what they imply for future Nell sessions. Nell's a complicated character with a lot of backstory and a range of emotions, neither of which I've really explored visually yet, so expect more of her in future galleries.
