Mr Wisp
patreon
The Rest: July 2023
🕑 Added 2023-07-01 04:00:06 +0000 UTCHello! Before you proceed, please take the time to note the TWO folders containing a combined 450+ 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.
Siraly Session #1 - Late February 2023
To start, I'd like to apologize for most of the Siraly images not being upscaled. That comes down to an oversight on my part: while the session was done before I decided that upscaling images would be the norm for these galleries, it is still possible to go back and upscale an image after the fact, which I could have done with these. I ended up not having the time or anlas to do that, so the only images from that gallery that are upscaled are those which I had upscaled during the session itself. Apologies for the inconsistency!
This session took place over a couple of days, but I've lumped it all together since the first half isn't particularly interesting to talk about: that would be the 'Siraly on the beach' set of images, which largely used lower strength settings. I think some of the tags here ended up detracting from the average image quality, though it's hard to say which. Initially, I wanted to have Siraly be eating cooked meat from a bone while walking on the beach, because girl likes her red meat. This ended up not translating at all for whatever reason, so I abandoned that focus fairly quickly. I was a little more stubborn about the tags 'skull necklace' and 'heart tattoo', which carried over most of the way into the second half of the session, but these tags never really got results, save for one instance where the 'heart tattoo' tag translated into something somewhat cool (see 'srl51 3' to 'srl51 5' in the second attached folder.) Overall though, this half of the session still managed to capture aspects of Siraly like her hair and figure pretty well, so I didn't linger on it for too long. One caveat I will note for this and the rest of the session is that Siraly is generally not a very happy or smiley person, and I did initially include tags like 'light frown' to represent this. But I ended up feeling bad for the character since she's had to put up with a lot of BS in canon, so I eventually caved and just let NovelAI depict her expression however it wanted, which resulted in some really warm and happy Siraly faces. Looking back, this was the right move: everyone smiles, especially when they're doing something they love, and Siraly really does love good food (even in a non-fetishy sense, she's a foodie through and through), so in my opinion the results still ended up being very believably in-character for her rather than just fanservicey. Also yes, she would be the type of person to take pictures of her food and post them on social media if that was a thing in Imperium. Somewhere, in some distant timeline, Siraly has a tiktok, and I'm a little sad not to live in that timeline.
That brings me to the second half of the session, aka the 'now she's on a bed and there's food' half. I absolutely love these images, and not just for the obvious reasons. I added background tags like 'hanging lantern', 'open window', 'dark room', and 'night sky' to try and give the vibe that Governor Tull's chilling in a private room either at an inn or her own residence, but this combination of tags ended up being greater than the sum of their parts, so much so that they gave the images a festive mukbang atmosphere that I'm not sure I could have communicated using those words directly. Siraly's also much, much more well-realized in this half of the session, by which I mean the average image quality compared to the first half is through the roof. There are still some duds and glitches that you can see in the attached folder, which are really just typical of NovelAI, especially at low resolutions: funky eyes and hands, floating bowls of food, and so on. But the image detail, poses, lighting, framing, expressions, and outfits are almost all above and beyond what I expect out of NovelAI in a typical session.
One thing I have noticed since this session is that base image really does matter when it comes to belly size specifically. I've used tags such as 'pregnant', 'big belly', obese', 'fat rolls' and so on across many sessions at this point, and these images were definitely the most responsive to those tags. In sessions where the base image doesn't have a belly at all, it's tough to get a belly as big as Siraly's to show up no matter how much you emphasize the tags, and you'll usually end up tanking the quality of the results before you get there. This has some favourable implications for drawing a quick belly on top of a base image in future sessions where I want to try generating some big belly shenanigans.
Finally, I must note the things in these images that I added by hand and that did not get generated by NovelAI: Siraly's eyebrow rings and chin piercing. I tend not to worry about getting these sorts of details to show up when generating images because of how infrequently NovelAI succeeds in adding them and how simple it is to just draw them in afterward. I did the same thing with the Nell session from last month, and I did a similar thing in the Selkie session from this month with quite a few of her features. Let's move to that session on that note.
Selkie Session #1 - Late April 2023
This session is an interesting case because I didn't remember it going very well, but when I looked back on it to compile this gallery I was surprised by how good some of the results looked. Most of these images feel very Selkie-ish to me, from her hair to her face to her outfit. Most surprisingly of all, these images actually succeeded in giving her a shark tail, which absolutely astounds me. I started this session pessimistically since I'd done similar experiments with unusual tails on humanoids before, with some dodgy results (no spoilers, sorry!) And to be clear, the shark tails in this session are far from perfect, to such a degree that I had to clean some of them up by hand afterward. But what I didn't remember working so well was their 'blueness' compared to the rest of Selkie. During the session, this disparity in coloration annoyed me a lot, but looking at the results now, I'm not sure if it was a coincidence or a secret genius move by NovelAI, because having Selkie's tail be bluer helps give a sense of underwater atmosphere and depth of field that I could never accomplish if I tried for it. Sure, in some images it looks odd or downright silly, but overall I really like the blue tail as a stylistic feature. I cannot overstate what a boon that is, because Selkie's tail is both one of her most important features and one of her most unintuitive to draw. I couldn't replicate it in her Vroid model, which I ended up using as the base image for this session; instead, I drew a shitty tail in using NovelAI's Inpaint feature, which allows you to edit base images within the app (very handy). But NovelAI got it, even if it didn't get it right all the time, and that honestly was the tipping point that made me happy enough with this session to feature it here.
HOWEVER, the tags I used to help make Selkie's tail happen, '{{{{{shark girl}}}}}' and '{{{shark tail}}}', also had their drawbacks, and I emphasize this both because it speaks to the NovelAI image generation experience as a whole, and also because it's kind of hilarious. If you've skimmed through the first attached folder already, you probably know what I'm talking about, but if not, let me warn you: there are some derpy-ass looking background sharks throughout these images, as well as some instances of multiple tails, limbs turning into tails, tails turning into background sharks, and so on. I think these instances are what stuck in my mind after finishing the session, which explains why I didn't remember it going very well. The reality is that this is probably as consistent a session as I could hope for with a character like Selkie, as not only did the stars align to make her tail look natural in the best of these images, it also still took a lot of enhancing and variation to get those images to a presentable state. In short, this session was a lot of work, and while it does encapsulate exactly what I wanted to achieve with this gallery (reasonably good and accurate AI interpretations of some unusual monstergirls), it also makes clear the limitations of the app. You can't go into it and expect every image, or even half the images you generate, to be good. If you have high expectations, it will take a lot of tweaking and fudging and luck (and cleaning up by hand afterward) to get the raw images to a state that satisfies those expectations. But if you're patient and can adjust your expectations over time, you can get really wicked results in the end! Just be prepared to meet some sharks along the way who are... pretty on the inside.
On the topic of adjusting expectations, I have to note the stuff I did by hand with these images: Selkie's head fins, eyes, gills, and shoulder tattoos. The head fins were done by using the '{pointy ears}' tag and editing those ears by hand afterward (this includes adding the little scratches on them, obviously). The eyes are something I have to cleanup by hand with every session, but in this case that basically just meant adding the red to them so that they were consistent. The gills and tattoos were entirely drawn by hand and weren't ever successfully generated by NovelAI, despite '{{{{{neck gills}}}}}' being a tag I tried emphasizing early on in the session. Similarly, features of Selkie such as her webbed feet and elbow fins just weren't going to happen, so I didn't bother much with them. I did include the tags 'elbow fins', 'webbed feet', and 'zora' throughout the entire session, on the off-chance that doing so would generate a couple of images that got close, but nothing came of it. Now, given what I've noted about the effectiveness of drawing in features like a belly and tail, there's an outside chance that drawing elbow fins and webbed feet onto the base image would have helped here, but I do get the impression that drawing features in gets less effective the smaller and more unusual the feature is. A shark tail may be pretty unusual, but elbow fins probably show up on a total of 2 images in NovelAI's database, so they remain too big an ask to be worth fretting over, and they're not a make-or-break feature of Selkie either. I was able to get the head fins and shark tail, so I'm calling that a victory overall.
Finally, let's talk about some of the fun underwater tags that made these images work! I used the tags 'underwater', 'submerged', '{swimming}', 'floating', 'seascape', '{{{pirate shipwreck}}}', 'light rays', '{{refraction}}', 'bubbles', and 'floating hair' throughout this session, with varied emphasis as I experimented with them. The MVPs here in my opinion were 'light rays', '{{refraction}}', and 'floating hair', as all of these tags put in tangible work that really pulled the resulting images together. I very rarely felt with this session that NovelAI was just depicting Selkie standing on land with a greenscreen ocean background behind her; instead, I felt reasonably sold on the idea that she really was underwater, which was a huge plus. I hadn't done much, if any, playing around with underwater settings prior to this session, so to have that aspect of these images turn out so well immediately helped me focus on trickier things like Selkie's tail. I can't say the 'floating hair' tag translated into every single image from this session, but when it did, those images typically ended up being among the best. I'm very excited to do some more experimenting with these sorts of tags in the future, be that in another Selkie session or with some other aquatic characters (cue the nervous Lamel noises).
Vespa Session #1 - Late June 2023
Here's the character inclusion I teased back when I announced this gallery's theme! This session took place between that announcement and the posting of this gallery, so it was brief and quick, but it went well enough that I decided it was worth including! The experiment here was to see if I could recreate someone as unusual as Vespa using NovelAI and a bit of editing by hand. I realized that starting with Vespa as Captain Blackhand would actually make this goal more achievable, since the trickiest parts of Vespa's design are almost entirely concealed by clothing here: her 'demonskin' forearms and lower legs, her little elbow spikes, her horns, and the majority of her bluish-purple skin. I'm not looking forward to finding out how these parts of her translate to NovelAI when using a base image and tags that are more typical of her usual attire (aka just a fucking apron), but not having to worry about them here helped narrow my focus down to a handful of things: namely, her ears, eyes, head shape, fangs, build, and hair and skin colour. I'll briefly address each of these things in order.
The ears were never going to be perfect since Vespa's ears have an extra pointy bit at the bottom, and while I could have added this pointy bit by hand after the fact, I didn't worry about it this time for the same reason I chose not to worry about Selkie's elbow fins: there were bigger fish to fry, lel. Vespa's eyes, on the other hand, I did worry about, and much time was spent trying desperately to communicate to NovelAI that this woman's eyes were blank. I did more research into NovelAI's tag system than I usually do, scanning several sites/guides that included eye tags, and eventually diving into Danbooru's tag groupings to get a better sense of what specific tags would make the most sense to NovelAI. Tags I tried included 'blank eyes', 'solid white eyes', 'no pupils', 'no irises', and 'empty eyes' (I was desperate at this point). I also added negative tags such as 'pupils', 'irises', and 'normal eyes'. Ultimately, there were signs that NovelAI was trying to respond to these tags, but it just never got over its apparent need to give Vespa irises of some sort. I feel kind of silly for trying so hard on this in retrospect, since all I really needed to do was go into a paint program and colour her eyes in with pure white, but it still would have been nice to get the tags to work on their own. We'll see what happens with future Vespa sessions. On the plus side, NovelAI nailed Vespa's eyeliner and general eye shape, so the Vespa vibe was there in some capacity even before I went in to edit anything by hand!
Vespa's head shape is a weird one. Usually I'm pretty lax about how my characters' head shapes translate to NovelAI since the way I draw heads is pretty janky. I do include tags like 'round face' and 'wide face' with characters such as Vespa and Ashley who have less typically anime-shaped heads, but I'm still not sure just how much work these tags put in overall. There were 'flashes of brilliance' during this session where NovelAI suddenly got Vespa's head shape right for a single image, but I attribute that to chance more than anything (see 'Vsp16 9' to 'Vsp16 11' in the first attached folder for an example). Head shape tags are something I research every time I work with a character like Ashley or Vespa, but so far I just haven't found much evidence that there are very influential tags in this category. I've yet to emphasize these tags very heavily in a session for fear that they'll result in some facial abominations, but I suppose that's a bullet I'll have to bite at some point if I want to really be thorough.
Fangs are a notoriously irritating thing to include as a tag, which is something I've gone into in previous months' notes. NovelAI just has a tendency to make characters with sharp teeth look like they crawled off the set of IT, which is cool for horror-themed images but less so for images of cute monstergirls. I had to deal with some of that fang jank during this session, but oddly enough it didn't seem as bad as it's been in previous sessions. I could hazard a guess as to why that is (I used the tag 'demon girl' throughout the session) but at the end of the day I'm still surprised, especially given how unorthodox Vespa is designwise. NovelAI struggled with a number of her features, as I'd expected going into the session, but her fangs weren't usually one of them. Imma just thank my lucky stars on that one and move on.
Vespa's build was pretty simple to accomplish, though I did try and fail to replicate the slight forward hunch she has going on in most of her reference images. NovelAI has demonstrated on multiple occasions that it has no problem with muscular women, and adding the tag 'broad shoulders' seemed to help with Vespa's overall commanding stature here. But I actually struggled to make Vespa less skinny here. Vespa isn't a slab of muscle by any means, but her waist isn't petite either, and that's something NovelAI had trouble comprehending. Entering the tag 'narrow waist' into the negative tags section helped a tiny bit with this, but it was still tough to overcome the app's innate bias toward giving muscular women Amazon-bodies. This is something I'm hoping to try and work around in future Vespa sessions, since representing characters' body types more accurately seems like it should be more doable than it ends up being sometimes, and that just drives me a little nuts.
Hair and skin colour are where I was the most consistently frustrated throughout this session. Too frequently NovelAI got one right and completely missed with the other, or it just got both wrong. Was I perhaps a bit too picky here given Vespa's unusual hair and skin colours? Yes, probably. Admittedly, NovelAI is going to struggle with any character a) whose hair and skin are the same broad colour but b) whose hair and skin aren't typical examples of that broad colour either. This was the case with Vespa: I spent a lot of time researching what sorts of tags I could use to communicate her hair colour specifically without just saying 'purple hair', and I landed on 'mauve hair', which did help somewhat with consistency. But I also experimented throughout with tags like 'light purple hair', combining 'white hair' and 'purple hair', and hybridizing tags (for example, 'white-purple hair' or 'white-mauve hair'). Based on some research I've done, adding dashes to tags does work sometimes with NovelAI to combine parts of these tags (https://naidb.miraheze.org/wiki/Prompt_Mixing_Guide), but in this case the results weren't consistent enough for me to stick with this method. In reality, Vespa's hair colour is unusual since it was chosen to complement an already unusual skin colour, so NovelAI's point of reference there is already way off from center. It doesn't help that it struggled with the skin colour too: purple-blue is how I'd describe Vespa's skin, but the problem with communicating that to NovelAI is the same problem with calling her hair 'white-mauve' in that a lot of the time both halves of that tag won't be represented equally. This meant that instead of Vespa's skin being purple-blue all the time, it was just purple or just blue a lot of the time instead. I played around with other methods of communicating Vespa's skin colour, such as including both 'purple skin' and 'blue skin' as tags but emphasizing or de-emphasizing either or both depending on the results. In the end, I had to be content with the images that portrayed Vespa's hair and skin reasonably accurately, which isn't all that different from other sessions featuring characters with non-white skintones. The Siraly session from this month certainly varied in skintone, so it wouldn't be too fair to expect NovelAI to be consistent with Vespa's... but it would be really nice to have more understanding of how to make skintone more consistent in both cases.
A couple more notes before I wrap up: 1) Vespa's fake scar was added by hand, not by NovelAI. 2) I tried using the tag 'night elf' at one point since I've realized recently that I probably unconsciously lifted Vespa's facial features from my several years spent playing WoW, but despite my expectations this tag didn't help her eyes and just ended up making the results look weirder (I think it was trying to give Vespa Illidan horns at one point, which is the point where I decided it was time to try a different approach). 3) I moved the Strength slider between 0.7 and 0.99 throughout this session, as sometimes keeping it lower can help capture more unusual details more accurately (aka Vespa's hair and skin colour, for my purposes here). The tradeoff there is that lower strength settings don't result in images of very high quality, so they're mostly just useful for seeing how close you can get to replicating the character's design while still getting a different and interesting result. On the other hand, keeping the strength at 0.99 has an annoying tendency to whitewash characters with nonwhite skin, since once the strength is at this point NovelAI's urge to make everything an anime girl becomes too powerful to control. Finally, any of the in-between points on the strength scale have the potential to embody the worst of the two extremes rather than the best, so navigating the pitfalls between 0.7 and 0.99 usually puts the quality of a given session at an overall disadvantage compared to sessions where you can get away with leaving it at 0.99 and going hog wild. I believe I've noted in previous months that the strength and noise sliders are the most difficult aspects of NovelAI for me to wrap my head around, but I'll mention it again here because it's still true! In many ways, these sliders are the most impactful settings when it comes to the appearance and quality of the results, so the fact that the 'right' balance between them remains so difficult to gauge throughout a lot of sessions is a constant source of pressure when using the app. But it also always makes things interesting and gives me hope that there's room for improvement and fine-tuning with every individual character, so in the end the best move is just to constantly be experimenting with them in cases such as Vespa's. As a whole, this was an eventful and challenging session, but the fact that I got some images out of it that were worth posting here makes me very happy and optimistic for future Vespa sessions.
