Posted on: 01/02/24 02:38AM
Yes, NovelAI. They have released a new model some 6 weeks ago and it's definitely the best I've seen so far.
Most importantly, it can now generate hands and feets now. Depending on their positioning and articulation, it might take several generations and some tweaking, but it is possible. It has inpainting, which means you can redo only the toes, if there's one extra or one missing. Hands are almost always at least passable now, often really good. Feet are a little more hit or miss, but with some practice and using suitable artist tags, you absolutely can get delicious foot fetish grade feet. Didn't manage to figure out gagging on foot, but the main things like footjobs, foot pov, etc. definitely work.
With the Opus subscription, you can generate an unlimited number of images with size of 832x1216 / 1024x1024 / 1216x832 or below, if you keep the default number of generation steps. This doesn't directly translate to "quality", the default can give you great images. If you put actual effort into the prompt and inpaint any minor glitches, if any, the results would often be tough to detect as generated.
It supports tags from Danbooru, very similar to Gelbooru, on top of SDXL's tags. It reacts strongly to the tags, but you can still use normal words - things like "slingshot ballbra" which aren't real tags will usually work. It can do all sorts of styles, either actual styles which I assume are from the base SDXL model that NovelAI then extends with focus on "anime", or styles based on artist tags. You can combine them and that works very well.
One of the great things about this is that you can "make" your favourite artist draw stuff they don't normally draw. Cutesexyrobutts doesn't draw futa for example, but this model can generate futa images that look so close to Cutesexyrobutts' style I wouldn't dare posting them anywhere. With famous artists, it's pretty much as if you had specific LORAs (a small dataset used to hyperfocus a Stable Diffusion model, which will then generate very similar images but nothing much else - some use it to mimic specific artists or for deepfakes), which are usually supported only when you run a model on your own PC. In other words, it's realy damn good at extracting styles even without any further tweaks.
All of the common fetishes should work. From toddlercon to matures, from lanky to walrus, chibi, western or anime, famous and even less famous franchizes, same with characters (anything with a few hundred images should work without any specification, along with translation into different styles - e.g. Misty (Pokémon) by Cutesexyrobutts, Misty by Renoir, Misty in style of kinetic pointillism - oh, that one came out really sexy actually) and scenes. It does get a little trickier when you combine several characters, especially if you try to precisely describe specific characters, as their description might start to mix. It takes some practice, sometimes a few repeated generations. All sorts of facial expressions, that part is nearly perfect and often feels handmade.
What does -not- work:
Swallowing part of vore. Guro only in limited extent and I haven't been able to properly generate something being cut/sawn off - amputees work, but mostly just the end product. Severed limbs are rather complicated to get, but not impossible. Getting a girl on a spit to roast seems very hard and rarely good, oven and grill somewhat work though. Some of the unusual penetrations like earfuck, urethral penetration (also insertion) or penis fisting are extremely hard, nosefuck seems nearly impossible. Urethral penetration / cock docking actually kinda works, but I can't get the glans right. Only tried it with futas though, perhaps it works better with men - actually, since I'm straight I can't vouch for quality of yaoi it produces.
You get the general idea, not really great for guro and rare fetishes you cannot describe using other tags. It can do "deepfakes" / nude filters of anime and cartoon content without a problem. It can do actual deepfakes too, sort of - assuming you're not a POS who seeks to bully or blackmail someone, but a well-mannered coomer who faps to their classmates and family members with respect and dignity in the privacy of their bedroom, well in that case it can turn ordinary photos into highly fappable content. Changing or removing clothes, if the pose isn't too crazy, works great. Overlaying details, like cum, is harder but doable. Realistic hands and feet are often trickier than cartoon ones though. You can use a realistic face and generate a cartoon body, it often works surprisingly well.
Ok I think that's enough. It's really great and offers endless fun. The results are, finally, so good I sometimes forget I'm not browsing handmade images.
Yeah and it can generate furry, it's pretty decent, but there's also an older Furry model specialized in just that (based on e621 and maybe Derpi I think), which will soon get an upgrade to match the newest anime model. This furry model has lower default (unlimited with Opus) image resolution, but it can produce better images and uses furry specific tags, so don't miss it just because it's in the Deprecated category. This furry model is also great at generating certain types of cartoon images and lolis.
Lastly, you can generate images at even bigger resolutions, but that costs a little extra. Based on subscription type, you get a number of Anlas / credits, and anything bigger than default costs a few, or a lot of those Anlas. You can buy more at any time - if this sounds like a mobile game crap to you, it's really not. Besides, you don't gain much by generating huge images, it's often better to generate several portrait-aspect images and stitch them together using inpainting or something. There's also an upscale function, which costs so little you could upscale ~1400 images per month with it if you had the Opus subscription. You can also buy Topaz Gigapixel for that during a sale or something btw.
Wait no, one more thing - it also includes a text generator, which is way better than ChatGPT when it comes to storytelling. It's not really a chat, it's collaborative writing mostly. You can use it like a chat too, of course. It also has amazing TextToSpeech, you could even use it just to read you text you paste there. I'm pretty sure they trained it on My Little Pony, among others - one of the default voices is pretty much Fluttershy. It's seed based, you can find seeds for other voices. I've found Twilight Sparkle, but I'm not sharing. I might give you the seed for Apple Bloom if you subscribe though.
P.S.: Whoever coded this forum, THANK YOU. Obviously, just as I've finished writing this I hit some WindowKey combination for Back, which I don't even know how to reproduce, and lost this entire essay. That might have been ugly. Luckily, the concept was saved.