Reposan a question from last thread that didn't get answered. Can I get a recc on decent, fast LLM models good for RP? I have no idea what to download and for what purpose. There's a enormous number of models and little in the way of documentation about them.

OP post is probably not great. Need to add guide for oogabooga & silly tavern for running LLMs locally, but I only see those guides on jewtube, not written. Renty's silly tavern guide did not appear to be for running locally. I reintroduced Request & Delivery anchors like back in the days as there are people who are better at these things & it's also nice to get ideas from others. Anyways, I played around with Ashley & style LoRAs so I'll dump some stuff. >>1020471 I liked using LLaMA2-13B-Tiefighter, but there's likely better stuff by now. Dunno.

Was hunting for a Budd Root style LoRA, but only found this Cavewoman character from that artist. Rather nice regardless.

>>1020469 Food inspired by your favorite vidya character

>>1020487 them titties are too small if it's trying to impersonate budd.

>>1020547 I tried to match size to another image generated using that LoRA & they didn't look small to me, but I've been generating a lot of loli so my perspective is probably skewed. I wanted to try figuring out how to gen multiple characters with regional prompter, but even with it installed & enabled it wouldn't appear as a drop-down menu like it's supposed to (ADetailer was installed the same way & appeared). In the end I wound up using the old ways with Photoshop & inpainting to get Hex Maniac & Gengar on the same image. Very time-consuming. I'm sure if the extensions to do it in the web UI worked it'd be just as much of a time sink regardless.

>>1020622 >>1020468 big pokemon tits

Was making a LoRA. Any other tag suggestions for this girl? >Elie_Lunar, red hair, ponytail, sidelocks, cape, yellow cape, jacket, collared jacket, raglan sleeves, dress, red dress, pantyhose, white pantyhose, boots, red boots,

What model for textgen do anons use? I'm using nous-hermes-2-mixtral-8x7b-sft.Q4_0 with Kobold, and it seems to have a strong habit for verbose run-on sentences that change topics wildly with a tendency for exponential increase in grandiosity and emotion.

>>1020469 hot looking babes with sexy dressing and big boobers.

>>1020686 female, girl, 1girl, cute, black eyes, short hair, long legs, smile, light smile

Used Photoshop & Inpainting to get things looking okay, then multiple SD Upscale images at different denoise settings blended together with Photoshop. With SD Upscale you can get better details at a higher denoise setting, but it always breaks anatomy so you have to do low denoise upscales too, then blend all the upscaled images together in Photoshop so you can get good details in most areas while not breaking anatomy in other areas. If there's one minor thing you want to improve you can crop out a 512x512 section of the large upscaled image, throw that into inpainting where you can be super-focused on generating one thing like eyes or fingers as everything else is cropped out, afterwards you just slap the improved 512x512 portion back into the upscaled image where you cropped it from. TL;DR AutismMix is a good name for the checkpoint I'm using since really pushing for good results is an autistic & time consuming endeavor.

Where last bread?

>>1021384 Read the OP: >>1020468 >Previous threads are moved to >>>/ais/ Previous thread: >>>/ais/4017 Previous previous thread: >>>/ais/3494

Also can anyone delete the background from >>1020686 ? Every method I've tried has proven broken.

>>1020468 ATTN: Please edit the OP post Remove everything after the divider & paste this instead (as it includes everything): Resources: https://pastebin.com/fLMeAMb1 Previous Thread: https://archive.is/4Ndbd >>1021402 After looking at previous threads, current OP is missing a lot of stuff right after the post cutoff as the post was copied without being expanded & the second long-ass post wasn't copy-pasted at all. Honestly, the OP posts have gotten way too long. To the point where you have two massive posts that have to be expanded, then a third small post after that. Putting ALL AI stuff in the same thread is convenient, but the OP post is insanely huge. It's at a point that you may as well just link ONE resource list so the OP post can be ONE post that is far quicker to make, less intimidating so anyone can do it & it doesn't become a game of waiting until someone else does it. If you don't like pastebin then suggest an alternative to de-clutter OP as I can't think of one.

>>1020490 >>1020470 Watermelon cake for THE STRONGEST 2hu.

>>1021417 That's more of a Photoshop request, but sure why not. >>1020470 Tried messing with this using IMG2IMG for fun, but really need a trained LoRA to keep those zig-zag patterns on everything.

>>1021641 Prompt/model/lora?

>>1021671 Some things were added later with Photoshop & Inpainting so were not in the prompt until I was adding them. I've excluded them for brevity as you're probably not gonna get that stuff normally. >Positive Prompt: ((score_9), score_8_up, score_7_up), score_6_up, score_5_up, score_4_up, 2 3 upper body, looking down, sitting at table, large plate on table with big cake, watermelon themed cake, 1girl, loli, detailed eyes, big head, detailed dark shadows, <lora:zy_Detailed_Backgrounds_v1:0.5>, detailed dining room background, winter, window, [pine trees, cloud], <lora:smr03_pony:0.7>, smr03, <lora:Cirno:0.7>, blue hair, ice wings, short sleeves, blue dress, short hair, hair bow, blue bow, puffy sleeves, white shirt, closed smile, <lora:Eyes_High_Definition:0.4>, eyeshd, <Negative Prompt: EasynegativeV2, PonyNegative, bad quality, bad_prompt_version2, score_6, score_5, score_4, watermark, mutation, muscular, lowres, low quality, extra limbs, text, signature, artist name, bad anatomy, poorly drawn, malformed, deformed, blurry, out of focus, noise, dust, artifacts, 3d render, (2girls, group, multiple_girls), depth of field, bokeh, (small head), bare arms, wings, teeth, nail polish, heart pupils, frilly collar, Model is AutismMix Confetti: https://civitai.com/models/288584?modelVersionId=324524 <lora:smr03_pony:0.7> is Samara03 Style LoRA: https://civitai.com/models/137595/samara03-style-pony15 <lora:Cirno:0.7> is Cirno - Touhou Project LoRA: https://civitai.com/models/514610/cirno-touhou-project-commission <lora:zy_Detailed_Backgrounds_v1:0.5> is Background Detail Enhancer LoRA: https://civitai.com/models/633524/background-detail-enhancer <lora:Eyes_High_Definition:0.4> is Eyes High Definition LoRA: https://civitai.com/models/670378/eyes-high-definition PonyNegative: https://civitai.com/models/332646/pony-pdxl-negative-embeddings EasynegativeV2: https://civitai.com/models/100191/ti-easynegativev2-textual-inversion-embedding bad_prompt_version2: https://civitai.com/models/55700/badprompt-negative-embedding Would likely be fine with just PonyNegative. The other negative embeddings are old. Regardless, I use too many LoRA & my prompts are probably too long as I'm nearly at the edge of breaking things sometimes.

>>1021665 Thanks. That one was a roadblock on making a LoRA for her.

And here's that LoRA I was making! https://civitai.com/models/799402/elie-ellie-lunar-magic-school

Neat, I was playing with a Midna LoRA. Wrangled the prompt for a while since it kept breaking. Eventually gave up & used a previous prompt I'd saved. Was easier to start from there than the broken shit I had going on. Also here's an earlier May K pic I didn't post as I was still trying to figure out the style I wanted for all the sisters when I did this here: >>>/ais/4463

Anyone know anything about upscaling video game textures? I have seen a few models on https://openmodeldb.info/?t=game-textures but I feel like that can't be the only place to find upscale models.

>>1022100 I don't remember where I got mine years ago, but I remember it had a crappy wikipedia layout. Maybe it was the site you linked & it had updated since. That site seems a lot better since it has filters & categories. I recall a manga upscaler that looked neat, but it didn't work for me. I got it working & tried to upscale an image that I can only find in a low resolution. I photoshopped the image a bit then put it through again to get halftone shading. Kinda hit & miss across the image, but certainly better than the low resolution source. I suspect it needs a cleaner image to do the crazy clean halftones I saw in demos.

>>1021959 Her head, hair, and helmet, and body art's good. Her body shape isn't impish enough.

>>1022100 Unless there's new development, AI upscaling video game textures looks shitty. Do them by hand, you'll need to seam by hand anyway.

>>1022202 It's hit & miss. >>1020469 More spooktober stuff. Doesn't have to be lewds.

Is there any hope of this ever not requiring a high-end GPU to run?

>>1022349 There's hope if you pay for cloud computing.

>>1020764 Midna mood.

>>1022161 Sounds neat. Could you do me a favor and either link it here, or try and upscale some pages from Candy Candy? The only scanlated version I could find looks like was last uploaded via dial-up. >>1022100 I use waifu2x generally but some pngs get fucked beyond repair if they have a lot of transparent areas.

>>1022349 Only in the sense that GPUs should theoretically keep progressing and eventually the current stuff will function on what in X years will be considered a low end GPU instead.

>>1022432 There's a lot of models on the page this anon linked: >>1022100 eula digimanga bw v1: https://openmodeldb.info/models/4x-eula-digimanga-bw-v1 <I used this older one because it worked on my Forge install. It attempts to clean lines & halftones, but I think if your source is really low resolution it'll struggle. eula digimanga bw v2 nc1: https://openmodeldb.info/models/4x-eula-digimanga-bw-v2-nc1 <This version 2 of digimanga is broken for me & unfortunately does nothing. eula digimanga MiA 65k: https://openmodeldb.info/models/4x-eula-digimanga-MiA-65k <Looks good & worked for me, but it is tailored for Made In Abyss manga or similar style. I only looked at ESRGAN upscalers because I know they can work with my install. There's other stuff I didn't try that's probably better. Just have to experiment more: https://openmodeldb.info/?t=manga >I use waifu2x generally but some pngs get fucked beyond repair if they have a lot of transparent areas I never tried doing transparency with SD, but I think I saw someone doing it with Comfy UI. Comfy is slower than Forge though from what I've heard. >>1022349 >tfw I had to sell my 4090 so I'm stuck using a 3080 Ti for Stable Diffusion 4090 was noticeably faster & so much VRAM. I miss the speed, but Forge is optimized a lot so it's pretty quick when it's not stuck dicking with memory management. >>1022437 True, but AI keeps progressing as well & may have higher requirements in future. Who knows. The main roadblock is private entities keeping things in-house so you don't have a major leap for locally run AI until something leaks like NAI did.

Heads up if you use the chink AI video maker! The moment you enter that site you will be automatically redirected to an english version of it, and that one requires you to sign in a (((google account))) to even make videos. There's a workaround I did where I refreshed the page & then quickly stopped it before the redirection. You will still be able to prompt your own videos & download them, but be warned that as soon as you leave the video window you will be dragged back into the english version.

>>1022466 Why not just use a VPN set to be in china?

Or just make a burner gmail?

>>1022477 >Why not just use a VPN set to be in china? But then you'll get a ding to your social credit score for making questionable loli content.

>>1020469 Requesting a regional prompting guide.

>>1022490 >>1020470 >regional prompting guide Saw one here: https://civitai.com/models/339604/how-to-generate-multiple-different-characters-mix-characters-andor-minimize-color-contamination-or-regional-prompt-adetailer-and-inpaint-or-my-workflow >>1022479 >make a burner gmail? How can you do that without a phone number? I've tried using temporary numbers to keep the account anon, but jewgle blocks using those numbers & won't create the account. Would be very useful to have an anon google account for downloading pirated games from gdrive & also since jewgle keeps trying to force a sign into youtube on my VPN & won't play videos. I've been using FreeTube, but I get the same sign-in message triggered on many VPN gateways,

Posted a new LoRA. https://civitai.com/models/822049/pratty-summon-night-swordcraft-story?modelVersionId=919243

>>1022497 I was able to make a google account without a phone number around December last year. If I had to take a guess as to why, it would be that people usually get new tech gifts for the holidays and it's better to get more people in the ecosystem even if it's missing what would be normally compulsory data.

>>1022497 I tried the examples from this guide for my last post and was hoping there was some other guide that was retard friendly floating around. Thanks anyway.

>>1022555 Ah. Guess I'll try around then. >>1022613 Farther than I got since Regional Prompter will install, but won't appear in Forge for me. Good news, everyone! I lamented in previous thread that there was no Pony LoRA for Kiki from Kiki's Delivery Service. Only 2 I found were from a Japanese McDonalds commercial redesign, not original design. I even found Pony LoRA's for other characters from Kiki's own movie. Well since then someone made one. Kiki LoRA for Pony: https://civitai.com/models/800687/kikis-delivery-service-kiki-pony Can't seem to grab tags from training data so just know that you at least need majokiki in the prompt to trigger it. Other helpful tags: dark brown hair, short hair, brown eyes, hairband, red hair bow, dark blue dress, baggy clothes, long sleeves,

>>1022750 Funny pumpkin on the left is me.

>>1022831 Someone turn it into an UOH

>>1022437 That just sounds depressing.

>>1022431 You should post your Midnas in Midna's thread too. >>799542

>>1022833 I had it as an UOH while inpainting, but it looked crappy so I changed it back. Here's something more UOH worthy. Kiki lifting her dress while I try to get Stable Diffusion to match her bloomers to the source material

>>1020622 What UI are you using? A1111? Comfy? Don't give up on Regional Promoter, it makes a big difference.

What should I tag her main garment as? Can't call it a dress because that would destroy the alt costume which is most definitely a dress. Also what kind of hat is in the outfit of pic 2? Razzy_SCS, dark-skinned female, yellow bandana, short hair, bangs, yellow ???, off shoulder, long sleeves, black tanktop, belt, buckle, bike shorts, shorts under dress, gloves, black socks, boots, Dress: ???, Yellow dress, sailor collar,

Also might as well ask for further tags for all the characters I'm prepping >Rifmonic_scs, cabbie hat, earflap hat, pink hat, twin braids, bangs, pink jacket, open jacket, fur-trimmed jacket, long sleeves, neck ribbon, black shirt, belt, skirt, white skirt, bag, sheath, weapon on back, black socks, pink shoes >Murno_SCS, blonde hair, hair ornament, forehead, halterneck, white dress, deatched sleeves, layered sleeves, long sleeves, black skirt, white garter straps, garter straps, black thighhighs, shoes thighhighs, white shoes, >Sanary_SCS, red hair, ponytail, neck ribbon, gray jacket, belt, sheath, weapon on back, black skirt, white boots >Rufeel_SCS, monster girl, ??? hat, animal ears, aqua hair, pom pom (clothing), strapless shirt, green shirt, puffy shorts, elbow gloves, fingerless gloves, tail, slippers >I'd make Lynn/Rinri too but I don't have remotely enough images >>1022903 Have you tried this LoRA? https://civitai.com/models/576402/bloomerssdxlpony?modelVersionId=642664 If it's not what you want, someone on CivitAI is taking free clothing requests and doing them fairly quick (I suspect this user is trying to get to #1 on the clothing leaderboard). I wouldn't mention Kiki, but "short bloomers"/pumpkin pants/パンプキンパンツ with appropriate example images from an image search might get you something. https://civitai.com/articles/7712/requestclothing-lora-requests

>>1022431 What's wrong with her legs, why are her legs so lumpy and fat? That Midna needs cellulite treatment.

>>1022956 In before someone posts Dagoth Ur's explanation.

>>1022976 If you're asking for it, I don't have it.

>>1022910 I'm using Forge. Regional Prompter installs & enables, but doesn't appear in the UI. I know the Regional Prompter files existed in the extensions folder as well, because I manually removed them when it didn't work, Multiple characters is still possible with a stupid amount of hours put into editing & inpainting. >>1020764 >>1020470 Okay.

>>1023135 I've been finally been blessed by the gods.

>>1022910 >>1023135 After some research I've found that Regional Prompter simply won't work with Forge. Even tried editing the python script with advice I found, but it always has errors & won't work on Forge at all. However, there's an alternative called Forge Couple: https://github.com/Haoming02/sd-forge-couple I installed that. It shows up in the interface & doesn't throw errors so far. I think it can only do two regions, but it's the same concept of separate prompts in the same image for different characters. I dunno if I can figure it out, but I can try it.

>>1023334 I'd say it's worth installing A1111 for it, I don't think the Forge improvements are that big of a deal, but being able to use regional prompter is. Even if you're only using two people, more regions helps more finely tune a scene. I don't think you actually need to move all your models/loras over, just set a symlink.

>>1022913 Training started. Even first epoch is pretty good. Third is even better.

Can I get any suggestions on LLMs to do kinkier stuff? A lot of them flounder around and don't know what to do for a lot of things.

>>1023135 Why is lesbianism so hot?

>>1023380 And published https://civitai.com/models/825000?modelVersionId=922566 >>1023431 Because yuri is the purest form of love.

>>1023369 I tried Forge Couple extension. Helps get a better start from TXT2IMG as I don't have to edit & inpaint the second character from something else. I started with both characters, yet still sunk hours into editing & inpainting regardless. I don't want to use A111 since Forge gens quicker & since I'm generating tons of images for nearly endless autistic changes any extra time spent waiting adds up. I can see how more regions would be useful to cover more than one area of where a character is in the composition.