>>6466 >I recently removed "bad_prompt_version2" from my negative prompts and my lolis' bodies looked way more Democrat supporting That's weird, bad_prompt and easynegative essentially do the same thing, the only difference is that easynegative does it with a LoRa while bad_prompt uses textual inversion, so I guess bad_prompt is a bit redundant, I just kept it there out of habit really. >the left pic is from my usual process and the right was me attempting the process you described. I like the left better actually, the right pic has cuter proportions but the transformation made the eyes and nose too realistic, which look a little freaky with anime proportions. When that happens I usually do a second inpaint of the eyes and nose with a more anime-like art style. >Also, you never explained the ultimate SD upscale script. Oh, sorry, for some reason I thought I had explained it, but it's not in the thread. Anyway, Ultimate SD Upscale is an extension, you can install it like any other extension using the github URL: https://github.com/Coyote-A/ultimate-upscale-for-automatic1111 Remember that scripts dropdown where you have the "SD upscale" option to increase the resolution of your pics? Well, this extension will install a new option in that same dropdown called "Ultimate SD upscale", which is basically a better upscaling script. The first pic are my current settings for upscaling with that script (you do upscaling from the img2img tab, by the way). If you use those settings you shouldn't even need Adverse Cleaner anymore because that generates a smoother upscale without all the graininess of my old methods. Also, here are some more lolis.

Any idea on how to fine tune a model using a particular image set?

>>6471 Are those really the settings you use for ultimate SD upscale? You don't use any of the sub upscalers? You keep the sampling steps at 28? You don't apply any kind of seam fixing? I tried upscaling an image by 3 like what you have here, with all these same settings, and the seams were very clearly visible. Pics unrelated; they were upscaled the old-fashioned way.

>>6493 Yep, pretty much. Most of the time, Seams fix makes no difference for me, and sometimes gives me worse results, so I don't bother with it. Same thing with the upscalers except for ESRGAN which makes the image noticeably grainier and it's the whole reason why I had to use Adverse Cleaner back when I was upscaling with ESRGAN. Here is a very quick example with zero retouching. Pic 1 is the original. Pic 2 is the upscale with the setting I posted. Pic 3 is with Seams fix. Pic 4 is with Seams fix and ESRGAN. Seams fix did literally nothing, and ESRGAN only made the image grainier. I suppose it bears to mention that I haven't update my Automatic1111 UI in a while. Checking the repository, I have over a thousand commits to pull. Looking at the log, I haven't updated since August 31st, so maybe that's making a difference? These pics should still have the embedded data in case you want to try to reproduce the results.

>>6493 >>6502 Welp, looks like 8ch automatically deletes embedded data from pics, so I uploaded them somewhere else: Pic 1: https://files.catbox.moe/487t07.png Pic 2: https://files.catbox.moe/ibtc2w.png Pic 3: https://files.catbox.moe/x791d6.png Pic 4: https://files.catbox.moe/e4c8ed.png

>>6502 >>6503 Alright, so I upscaled 478t07.png and ended up with pretty much the exact same results as you; not a seam in sight. And you were right about ESRGAN, it was entirely responsible for the graininess. I typically use a lot of tags and LORAs when generating an image, so I think probably one or more of the LORAs I normally use is responsible for the noticeable seams. Anyway, thanks for all the help, you're awesome! And hopefully my own experimentation gives you some insight. I've been keeping up with SD's updates recently and I don't think you're missing anything.

>>6502 >>6503 >>6505 Your generations definitely have brighter colors than mine. Do you know why that is?

>>6506 Oh, that's because I forgot to reenable color correction in the im2img settings. I sometimes disable that when changing art styles to avoid some weird color artifacts, but since the option is hidden away in the settings, sometimes I forget to reenable it, which tends to slightly increase saturation each time you run the pic through img2img. The colors in your upscales are actually more accurate to the original, mine got oversaturated because of that option I forgot to check.

>>6509 Oh, I see.

Old gens but never had the opportunity to share them before

Nsfw loli

Some new pics. In case you're wondering about the last pic, I used this: https://huggingface.co/monster-labs/control_v1p_sd15_qrcode_monster

I wasn't entirely happy with the clothed versions of the elf loli, so I gave it another shot.

>>6455 Glad to see you're still hammering away at it and refining your AI tard wrangling. Can you make a nopan variant of that second suspenders loli? Also, that Lina Inverse loli might have something weird going on with her ass. Her right (viewer's left) ass cheek that's visible from the front looks like it's going inside of her pussy, sort of hanging out of it. >>6466 >Democrat supporting Lmao, wordfilters.

>>6547 Have a variant, a fix and a new pic.

>>6553 Saintly.

>>6328 Good in XL models but going beyond 1024x1024 ratios can introduce extra toes.

Yo, Uotism? You alive? It's been over 3 months, and I think that's a new record.

>>6709 Sorry about that, I'm alive, just very lazy and unmotivated. The new XL models have been giving me a hard time because of the censorship that resulted from the anti-AI push. Now that loliDiffusion XL finally released, one of the censorship issues has been solved, but my biggest problem now is that all of the art style tags that I've built over time pretty much stopped working completely in the XL models because people don't train models on artist names anymore. I can generate generic anime loli just fine, but I feel like that's kinda missing the point of using AI. Like I said before, the thing I like about AI is doing stuff that would be extremely difficult or impossible to do by hand, so if I can't get the art styles that I like, I don't feel like I have much to offer. Also, a new wave of people have started using AI to make lewd art and I feel like I have fallen behind by quite a lot. The solution I've been marinating in my head was to use the art style switch I talk about here >>6455 Basically I make a base pic in a generic anime art style and then switch to the old SD1.5 based models to change the art style. This is very annoying to do because loading and unloading models takes a long time, prompting in the new XL models is very different from the old ones and all of the LoRas and VAEs from the SD1.5 models don't work in the XL ones, so to switch back and forth because I have to change everything I have loaded in addition to the model itself. Your post made me finally get off my ass and try doing exactly that, here are the results. I like that I'm able to make more complex pics than just 1girl posing, but I'll have to find some way to make the process less annoying, because just these two pics basically took a whole day to make between all of the model switching, inpainting, and even a lot of manual painting with a mouse.

>>6730 Sad to see shit's got all fucked up. You should post in the /v/ AI thread since the issues you encounter effect more than just loli and could be useful info for everyone. I finally got a real computer and almost thought I could make something half-decent to satisfy my need for loli versions of characters, but I fucked something up along the way and couldn't figure out how to fix it. Here's what I ended up with, and here's my posts on the subject >>>/ais/4003 >>>/ais/4004 >>>/ais/4007 >>>/ais/4009 >>>/ais/4010 >>>/ais/4011 >>>/ais/4012 >>>/ais/4016 >>>/v/988620 I still have over 600GB of loli models, but I don't know which ones I actually need, nor do I want to trial and error my way though that massive mess when I'm not even competent to determine if I'm using a model right. I gave up, moved on to text generation, and gave up on that too when I started encountering a persistent issue with the AI dropping the second person perspective.

>>6730 >people don't train models on artist names anymore. They got tired of artists claiming copyright on their art style (something they've never owned intellectual rights to). This way it is also harder to prompt the AI into recreating one of those artists actual works: trying to force the AI into copyright infringement (the IP version of posting CP to report it).

Here are some more pics using the art style switch method. They are certainly more laborious to make compared to the old way if doing it, but I'm decently happy with the results. >>6731 Right now I'm doing a lot of model switching, I start with loliDiffusionV1.1.0_PDXL_Lx7 for the base pic, then I use Based64mix-V3 in img2img to switch art styles, this takes a couple of tries and mashing parts of the results together in krita, then I do the upscaling also with Based64 (sometimes I vary the model depending on what the composition of the pic is). When I have a high resolution pic, I fix the defects with inpainting using mostly Based64, but sometimes I switch to XL models if I have to fix something complicated because they are smarter in recognizing things like hands or specific objects, so for inpainting I'm constantly switching between Based64, tunixStylizedPony and dreamweaverPony25DMix. I also like to do the hair in a hyper detailed style like in the 4th pic, to do that I switch to cardosAnimated_v10, load the add_detail LoRa at 2 strength and select the hair for inpainting, depending on the results I might do a second pass with a different model to blend it in better. After all of that is done I do some color balancing and contrast tweaking in krita to get those sexy reddish skin tones that I like from Boris Vallejo's art style. I think what most people miss about using AI is how powerful a tool inpainting is. It basically allows you grab a pic and flawlessly blend anything you want into it. You can fix almost anything with inpaitning, so even if you start from a shitty base pic, but you like the composition, you can gradually turn that shitty pic into something amazing. >>6739 I remember seeing some of that a while back, some guy was using img2img to slightly alter a pic he made and then posted the comparison to twitter claiming that the AI stole his pic.

I think it's time for a new thread. Every time I try to reload the page, I get the "could not reload captcha error" and most of the files fail to load. I even sometimes get "429 slow down asshole" just from this thread. >>6745 >I think what most people miss about using AI is how powerful a tool inpainting is. It basically allows you grab a pic and flawlessly blend anything you want into it. You can fix almost anything with inpaitning That's what it seemed like when I used to easily fix the arm of the redhead in >>6731 , but when I got around to using inpainting on 2B, nothing seemed to work anymore, and it was a struggle just to do small fixes. The final pic is missing some removed black lines from the hair that I forgot to save, and those took way to long.

I think I'm getting the hang of this, here are some more pics with the new method, again trying for more complex scenes. >>6750 I'm not having issues loading the thread, but yeah it's gotten pretty big, I'll have to ask the mods if I can make a new thread while this one still hasn't hit the post limit though. As for inpainting, if the AI keeps misinterpreting what's in the target area, it's usually because it's lacking the context of the bigger pic, see the tip about "context points" here >>5389 Also, what's your denoising strength when inpainting? And what's your inpaint area setting? Because that should almost always be on "only masked".

>>6752 >I'll have to ask the mods if I can make a new thread while this one still hasn't hit the post limit though. If you want to make it just make it. Or hell, claim the board yourself. /loli/ is abandoned because the BO just squats the place until someone tries to claim it, at which point, he's contacted by the admins to see if he's still alive, moderates for maybe a couple weeks at most, and then fucks off again.

>>6752 Damn, you weren't kidding about that scene complexity. This is some good shit.

NEW THREAD >>6756 >>6756 >>6756 >>6756 >>6756 >>6756 >>6756 NEW THREAD