Mr Wisp
patreon
The Rest: December 2023
🕑 Added 2023-12-01 05:00:07 +0000 UTCHello! Before you proceed, please take the time to note the SIX folders containing a combined 750+ AI-generated images attached to the end of this post! There are six 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. Also, these images come from four separate sessions, each of which I'll describe individually, in chronological order. I'll 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.
Lani Session #4 - Mid October 2023
Aka: 'Lani Sykes, Pie Glutton'
These sessions are all very recent because they pick up where last month's gallery left off, with this specific session having happened shortly after session #3. Thematically and promptwise, this session is very similar to the last, but I changed up the pose and some of the food-related tags to try and make it more pie-centric. My mental image when starting this session was Lani in a room full of pie- an image I sadly did not achieve, though not for lack of trying. I described in last month's post how increasing prompt guidance is a double-edged sword when it comes to 'pie', due to Danbooru's interest in a kind of pie that isn't baked nor Lani-friendly. This issue carried over into this session and made me hesitate to have NovelAI follow the prompt more directly. Additionally, I felt having a room full of pie would require more creativity on NovelAI's part, which upping the prompt guidance value wouldn't help with. Instead, I focused on experimenting with more specific pie-related tags that I hoped would communicate more clearly that I wanted to see baked goods and not that other thing. 'Incoming food', 'incoming pie', 'too much food', 'too much pie', 'oversized food', and 'oversized pie' were used to emphasize the presence of pie-as-food (and to try and achieve the fabled 'room full of pie' effect), but I dropped most of these tags eventually since they never seemed to affect the results in any meaningful way-- with one exception. Image ln228, a variation of which is included in the main gallery, is the closest I got to the room full of pie, but it was the exception rather than the norm, and I had a tough time replicating the effect. I maybe could have used that image's seed and tried to add more and more pie to it, but that idea didn't occur to me at the time, and even now I'm unsure if it would have helped or just resulted in more of the same image.
'Apple pie' was used at some points for a similar purpose, as I thought that specifying the type of pie would make pie as a whole more likely to appear. There's evidence in some of the results which indicates this tag may have helped a bit, but unfortunately it often had the effect of replacing the pies with apples, which you can see in one of the main gallery images. I rolled with this for a little while because I thought it was funny, but eventually it got irritating and I dropped the 'apple pie' tag entirely. Overall, I can't exactly complain about the amount of pie present in this session's results, as there were plenty of instances of Lani holding one or two pies, which was an improvement over the previous session. But ultimately the need to keep fiddling with the food-related tags became tiresome, and I moved away from that specific sort of experimentation in later sessions.
Something worth mentioning about this session specifically (that hasn't happened in any other session before or since), is that my GPU glitched out near that end and messed with the image generation process, resulting in a series of images where the lighting and colouration were way off what they should have been. You can see these results in folders 5 and 9, as well as in a couple of main gallery images: they're the images where everything is dark except for Lani's hair, eyes, and teeth. I have no idea what about my GPU spazzing out caused this change, but I ended up liking the vibe of these images, so again, I rolled with it. The resultant 'Lani Night Mode' mini-session made for a nice cap-off to this session and the previous one by making it feel more like a progression from day into night, giving some sense of how much time Lani would spend indulging in her pie obsession if given the chance. If the night mode images come off as slightly eerie, then all the more fitting, I say.
Lani Session #5 - Mid October 2023
Aka: 'Lani Sykes, Indulgent Beauty'
I was on a roll with Lani by this point and felt I'd gotten enough solid results that I could get away with getting really silly for the sake of experimentation. This goal this time was simple: make Lani as fat as possible. The base image I gave NovelAI for this session is the attached panel from the 'Lani and the King Pie' comic featuring a very, very absurdly large Lani (file name 'lncrwn'.) I tried making up tags such as 'immobile' and 'impossible belly', hoping these would somehow help convey the sheer size of Lani in this image, and I put all the belly-related tags at the head of the prompt to maximize the odds of them being recognized. I kept the background tags relatively simple as well, or at least I intended to: I figured having her sit on a patch of grass with some blue sky and a fence behind her would allow NovelAI to focus more on Lani rather than worry too much about the background, but I think including the tags 'wooden fence' and 'against fence' was a mistake in this regard. In hindsight, I would have airbrushed out the fence and saved myself some headache, but then again, the average image quality this session was so mid that I don't think the janky fencing was the primary issue. In short, NovelAI could not process Lani's size (particularly that of her belly, which I don't blame it for, really.) There are a few results from this session where Lani is very, very large, but none of them were quite as ridiculous as the base image, and those that came close were almost always very flawed images for multiple reasons. There are some very unflattering images of Lani in attached folder #5 which made this session feel more pinup-y than the rest, though in all fairness I don't think my prompt or base image helped with this at all; I'm not sure what I expected NovelAI to do with tags like '[spread legs]' and 'black panties', so that's on me.
I actually started this session off with a high prompt guidance value (14.5 to be exact), to see if that would help NovelAI comprehend Lani's abnormal belly size at all. Doing so may have helped a little bit, but not nearly as much as I'd hoped, and it wasn't worth the resultant drop in image quality. Lowering the prompt guidance to 6 (and removing the 'spread legs tag') immediately resulted in better quality images without compromising Lani's size at all. In fact, I was eventually able to get results such as images 361, 372, and 394 in folder 6, which are just about the biggest Lani's gotten in any session without the image quality being complete garbage. But these results weren't consistent, and I ended up accepting that Lani didn't need to be as gigantic as the base image if that meant higher quality results. In other words, I gave up on trying to make her a blob, which was just fine in my mind. I've moved away from drawing characters at those sizes anyway since it's not my favourite subject matter, and there just isn't enough of that sort of content on Danbooru for me to expect NovelAI to be able to replicate it at this point (this is, of course, speaking in terms of the 'NAI Diffusion Anime V1 (Full)' model, which is what I used for most of these sessions. I'm unsure if newer models are any better in this department, as I've yet to test them out properly.)
Lani Session #6 - Mid October 2023
Aka: 'Lani Sykes vs. Your Unwanted Pies'
This was another experimental session geared toward making Lani as big as possible, but this time I approached it from a more stuffing rather than weight gain-centric perspective. The base image I used here was the original stuffing image I drew of Lani several years ago (file name 'lndnrght'.) I was curious if NovelAI would find it easier to replicate a belly of this size if it was rounder and more spherical rather than a blobby mass like the previous base image. The short answer was: yes, but the average quality of the results still wasn't great. There's a reason why only two images from this session made it into the main gallery: I kept the img2img strength between 7 and 8.5 for much of this session as it resulted in larger bellies, but at the cost of these images basically being bootleg reproductions of the base image. Higher strength settings resulted in familiar glitches that I encounter every time I try to push the belly size with NovelAI: those being the emergence of mysterious beanbag chairs and other abstract objects where the belly should be. Occasionally, these objects are close enough in appearance to an actual belly that I can almost get away with using them, but the major roadblock there is the accompanying lack of colour that usually afflicts these not-quite-belly shapes. This train of thought is what prompted me midway through this session to try something a little different: if I added the tags '{{{black and white}}}', 'comic screen tone', '{{{{{{{{grey scale}}}}}}}}', '{{{line art}}}', '{{{{{masterpiece illustration}}}}}', could I use a manga-esque colour scheme to mask the fact that the belly was discoloured? Again, the short answer was yes, and I honestly think the idea's worth another session's worth of experimentation at least. But the bigger problem that cut this session short was the aspect ratio of the base image, which is a ratio that NovelAI isn't as familiar with compared to the typical portrait dimensions I work with for most sessions. I tried eventually changing the aspect ratio to those more typical dimensions, but doing so of course skewed the base image and reduced the likelihood of getting natural looking results out of it. No matter which dimensions I used in this case, the results were going to look weird in some way, so I concluded that it wasn't worth the continued effort. If I try the manga-style experiment again in a future session, I'll use a base image that already has portrait-like dimensions so that I can avoid unnecessary drops in image quality. With all that said, this session definitely produced some intriguing and occasionally unnerving results, and the black and white screentone played toward this more unsettling quality in ways that felt strangely fitting of Lani, so I am genuinely curious to give it another shot in the future.
Lani Session #7 - Late October 2023
Aka: 'Lani Sykes, off the Wagon'
I'll start off with a cursory note on the two images in the main gallery featuring burp onomatopoeia: these were added by hand, as it's almost never worth it to try and get burp effects through NovelAI. My approach to it is to add them in myself when I see an image that looks like it deserves them, as they don't take much time to draw and (I find) don't distract much from the rest of the image. That's why Lani sometimes be burpin'. Thank you for coming to my TED Talk.
This is the latest session I've done to date, and the only one to use NovelAI's new 'NAI Diffusion Anime V2' model (which I'm just now discovering is itself outdated, as 'NAI Diffusion Anime V3' is now a thing.) From this session alone, I get the very strong impression that 'Anime V2' handles chubby girls much better than V1, as this was one of the most enjoyable and interesting sessions I've had with NovelAI. I returned to the tried and true subject matter of 'Lani eating pie in bed' for this session, but with a few new tags sprinkled throughout: 'coloured shadow' and 'fisheye' are two tags I used for the first time with this session, and both put in work, though I got what I wanted from the 'fisheye' tag fairly quickly and thus cut it from the rest of the session. I added in the '{{realistic}}' tag midway through for good measure. Otherwise, there wasn't a ton of experimentation with this session, as I was too impressed by the new Diffusion model to mind. The strength was set between 0.9 and 0.99 throughout, and the prompt guidance value was kept at 8 (get it, because that's what Lani did to all those pies, hue hue hue hue.) I see this session as a proper counterpart to session #3 from last month- it's sunset now, and Lani looks on average slightly thiccer, which I credit to Anime V2 handling fat more smoothly than V1. The variety of belly shapes and sizes throughout this session was impressive, and the only reason I didn't do more sessions following this one was because I had to focus on some big real life events that limited my ability to use the internet for a while. I'm curious if the new models will allow for larger Lanis now, as that's not something I pushed particularly hard for with this session (at least not compared to the previous two sessions.) The prospect of doing yet another Lani session feels a little silly at this point, as I believe I've spent more time on her in NovelAI than any other character, but my continued interest in seeing if I can represent her maximum sprite size in Axugaem2 through these sessions impels me to give it at least one shot with Anime V3. Expect to see the results of that within the next couple of months.
This talk of the new Diffusion models brings me to somewhat of a bittersweet conclusion for this month's notes, as well as a major announcement (I'll be making a proper post for said announcement soon, but I think it's worth including some of my thoughts behind it here since they relate to my experience with the AI image generating process): although I should feel encouraged by these advances in AI art, I honestly feel a bit of the opposite effect. It's not that I think the new models are bad or uninteresting at all! Rather, the turnoff for me has been that they appear to be getting too good at mimicking the style of 'mainstream' fetish art now (or I'm just starting to recognize stylistic flourishes, techniques and so on that I've seen in real fetish art drawn by real people.) This is just about the biggest red flag for me, as I don't under any circumstances want to start ripping off specific artists with these galleries. Putting aside the fact that AI art is a sore topic for a lot of artists, and I cannot blame anyone for feeling offended, threatened, annoyed, hurt, or ripped-off when their work is being mimicked more and more closely by AI models, I don't think it's very interesting or productive for me to continue spending so much time with this project if the results are going to become more and more like existing works of art by those in fetish communities.
I understand this is very hypocritical of me, as AI art always draws from existing works of art by someone, so why should it make a difference if I feel like I'm ripping off bigger artists?
I don't have an excuse there, and I think that's what's really putting me off AI art at this point. I've always felt that the distance between NovelAI's image database and specific fetish communities kept the results fresher: what I mean is, Danbooru doesn't have a ton of content tailored toward weight gain, belly expansion and other such fetishes (though there's more of that sort of content every day, as is the law of the internet,) and what it does have tends to approach the subject matter from a different stylistic perspective than, say, DeviantArt, meaning that when I use NovelAI, I have to work a little harder to bridge the stylistic gaps between what NovelAI and I consider to be a 'big belly' or 'feedee' or whathaveyou. Is that because Danbooru features content from primarily Japanese artists? Probably! Has my ignorance of and distance from Japanese notions of fetish art been what's kept these galleries feeling 'fresh' to me? More probably than I'd like! I still don't know Danbooru and other such communities well enough to really recognize when a result I get using NovelAI is drawing really heavily from a specific artist that I just don't know because of cultural distance, so I can't say if that's what's happening. Ignorance in this case is really bliss, but having worked with NovelAI for over a year, I'm at least a little less ignorant of these communities than I was when I began, and the 'magic' of AI art is certainly wearing a little thin on me.
I've mentioned in previous months' notes that I may move on from the Re-AImagined project soon, and in light of these anxieties about who AI art draws from and how heavily, I feel ready to refocus on drawing my own stuff again. Doing these galleries has been a lot of fun, and I've learned a lot of useful things about art and AI from the process, but in terms of my personal growth that's come from experimenting with AI art, I think I've reached a plateau. Thus, I'd like to officially give a one month notice on these galleries: January 2024 will be the last of them, at least for the foreseeable future. There are a couple of other factors influencing this decision: namely, I have the space and the means to spend more time drawing again, and I want to continue making Axugaem2 my primary artistic focus. But my slight disillusionment with AI art has been the tipping point. Again, if I haven't stated this clearly enough: I don't think AI art is bad or evil, and I don't think anyone is a bad person for using or experimenting with it. My anxieties with it are largely personal and opinion-based. As an artist who can just draw my own fetish art if I want to, my interest in AI has always lain with the idea of trying to make new and different images, which I believe/hope has been the case sufficiently enough to justify this whole venture. I don't believe I'm special in that regard either, and I encourage anyone working with AI art to continue doing so if they feel what they're doing is important. Just also keep an open mind and recognize the debts and limitations of the form.
Rather than end on a low note, I'd like to say thank you to everyone who has enjoyed and supported the Re-AImagined galleries over the past year! Your interest in the project has kept me going and pushed me to try out some really fun and interesting ideas, so I truly appreciate it! As mentioned, I'll be doing one last gallery for January, and I'd like to take direct suggestions from you regarding what you'd like to see in it (I'll be repeating this information in a separate announcement post, so don't worry about commenting on this one.) I'll be going back over my notes from this year to try and put everything I've learned together for this final gallery, so we can give this project a big sendoff! In terms of sessions I've yet to post here (of which there are a few), those will either make an appearance in next month's gallery or appear as bonus content at some point in the future. I won't put the possibility of eventually doing another gallery or two out of the question either; depending on what the future has in store for AI art, I can see myself revisiting it somewhere down the road. For now though, count on next month's gallery being the grand finale!
