Previous thread: >>4494776 >LOCAL UIreForge: https://github.com/Panchovix/stable-diffusion-webui-reForgeComfy: https://github.com/comfyanonymous/ComfyUI >LOCAL MODELSNoob: https://civitai.com/models/833294?modelVersionId=1116447 | https://civitai.com/models/833294?modelVersionId=1190596Illustrious: https://civitai.com/models/795765 >RESOURCESWiki: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki | https://comfyanonymous.github.io/ComfyUI_examples | https://openart.ai/workflows/templatesTraining: https://rentry.org/59xed3 | https://github.com/derrian-distro/LoRA_Easy_Training_Scripts | https://github.com/bmaltais/kohya_ss | https://github.com/Nerogar/OneTrainerTags: https://danbooru.donmai.us/wiki_pages/tag_groups | https://danbooru.donmai.us/related_tagControlNet: https://rentry.org/dummycontrolnet | https://civitai.com/models/136070IOPaint (LamaCleaner): https://www.iopaint.com/installUpscaler models: https://openmodeldb.infoBooru: https://aibooru.online4chanX Catbox/NAI prompt userscript: https://rentry.org/hdgcbIllustrious-related: https://rentry.org/illustrious_loras_n_stuffUseful Nodes/Extensions: https://rentry.org/8csaevw5 Text Generation: >LOCALhttps://rentry.org/lmg-spoonfeed-guidehttps://rentry.org/LocalModelsLinkshttps://rentry.org/local_llm_glossaryTextgen backend: https://github.com/oobabooga/text-generation-webuiWorldbuilding frontend: https://github.com/SillyTavern/SillyTavern >MODELSGeneral leaderboard: https://hf.co/spaces/HuggingFaceH4/open_llm_leaderboardUncencored leaderboard: https://huggingface.co/spaces/DontPlanToEnd/UGI-Leaderboard
Not sure why a new thread was not made let the fun continue
>>4533322
>>4533323
>>4533324
>>4533325
>>4533326I had to take advantage of that while it was still usable; there are problems with their models and NAI doesn't recognize them either.Now with these two, I used the very brief trial that Banana Pro gave me, but I thought I would have at least 2 free tasks per day, but that was not the case and I was left with nothing.
someone last thread said copyrighted characters don't work with banana pro anymore, but during the downtime of /udg/ I found out it still worksand found a new hobby, replacing canon het screenshots into superior fanon
>>4533358It's still far more censored now for suggestive stuff.
How about shipping characters with same va?
>>4533596
>>4533597
>>4533599
>>4533600
>>4533601
>>4533602
>>4533603
>>4533467nice pose, did you prompt or CN it?
Haven't been here in a bit. Anyone started using Anima, yet? Being able to use plain text in prompting along with tags is pretty useful. Messed up a bit on the example prompt. It was supposed to be "has blonde hair and closed eyes" but hopefully even this very simple prompt still shows the ability to assign actions to characters well enough.
Greetings from /u/lit/
Pop Chaser came out 41 years ago today.
>>4537328Too raw and alpha. I wish z-image team did a training on gelbooru database like they promised before.
can I /r/ some gunsmith cats minnie may?
>>4538252
Can I request Kei from Dirty Pair and Kaori from City Hunter kissing?
>>4538325
Is there a tool where I can put character images together? It has to be free; I don't mind a limited number of attempts per day. The problem is that my character selection isn't popular, and I've had trouble with Lora for yuri.
>>4538745Nano Banana Pro, but supposedly its quality for using references has diminished lately.
>>4538750I tried to use it because they told me I had 3 chances per day, but it only gave me 2 and they didn't renew. Did I do something wrong or did I choose the wrong page?
>>4538751NTA, google gemini is the standard for nano banana pro(or rather nano banana 2 lately). Normally it can do 3 since the last time I used it, however it's true that failed attempts may count as (hidden)credits usedWhen they switched from pro to 2, I noticed I could render a lot more now, though that may have been just for a few days, haven't checked it out lately
>>4538745Don't use services, use local models only. Qwen edit is the best and completely uncensored. Also flux klein is a great option too, it requires even less vram.
>>4538256thank you kindly
>>4540126and two that didnt quite come out right
>>4540128lol hands>>4538175understandable, they're at least releasing updates and not abandoning it like some other models so hopefully in a couple months it'll be a lot more stablei also do wonder about what happened to that z-image train
>>4540449
>>4540451
>>4540453
>>4540454
>>4540455
Which UI do I migrate to? I tried forge but inpaining didn't work nearly as good. I need good UI for creative work.
>>4544766cumfyui duh
>>4544859It's a misnomer, there is nothing comfy about that UI. My finger hurts just from constant panning and zooming I have to do to use it, I can't imagine actually using it for anything creative like inpainting.
>>4545868For inpainting, you're supposed to use Krita (an open-source Photoshop analogue) with its ComfyUI plugin.
>>4546450you can inpaint in comfy itself it's mask editor has improved since the past and i actually found the krita extension to be clunky to use though it has been a whilei personally use paint.net for any edits i need to do then throw it into comfyui with the impact pack nodes for the actual inpaint processstill more clunky then forge (which's inpainting should be the same as A1111) though
Huh, I can't seem to gen anymore. All the links I get are 403.
ubel corrupting fern is a nice idea
>>4534695How ? Can't find a good website or local llm to do that
>>4540449>>4540451>>4540453>>4540454>>4540455moar plslove size difference so much even better if younger girl with smaller older girlthe yun yun ones is something that would happen
>>4549296Here is some moaarAlso it's really sad that current models can't do anything more complex without spending endless hours praying to the RND goddess.
>>4549479
>>4549294It's just Grok at his best
>>4549484
>>4549484thanks anon>>4549923hhnnggg wish she ends up even taller than huge dwarf guygreat work anon
want to be a tiny Erza so Wendy can have her way with menot enough younger girl with smaller older girl/woman content
>got into genning just a last week>spent a whole day cycling Loras and artist styles and tweaking small details for a single sceneI am genuinely amazed, this is truly awesomeBit of a shame this general's so slow, not that it's always necessarily a bad thing (the other boards I've checked out are filled to the brim with schizos and drama queens) but still
>>4551437there are hardly any request, at least not much spammer either
>>4551734>>there are hardly any requestCan I request an image?
>>4551738sure
i stole this from /b/
>>4552192/b/ AI threads are an absolute cesspool of degeneracy, but there are some hidden good /u/ gems from time to time
Amazing! I can use gemini to substituted a male character from a manga and turn into yurihttps://imgur.com/a/Zkoixfi
>>4552353Oh my go, so many posibilties
>>4552354Oh my! is really good!
>>4552353>>4552354>>4552355Damn. The future is now.
Well, this is the best for now for a more lewd scene
Last one, will tried a bigger project and convert a lot of this manga into Yuri
>>4552353yeah, I have used it on anime screenshots before, but I noticed I don't know that many shows that have onscreen romantic hetero scenes I would want to convert haha.though honestly I haven't even considered doing it for manga
>>4551859I would like Yuri with the two of them. As I mentioned at some point, using Lora with the two of them apparently doesn't work because of their similar designs and the characters tend to blend together.
On the other hand, I didn't expect chatGPT to do such a good job with an image of the two of them together (they are Sushiko Maki and Finger Chikako from Yugioh Sevens, the missing couple from the food club in that series)
>>4552511litterbox is broken, had more ...in general forge couple works great for seperate loras, since anima didnt have those chars well enough i used illu model and 2 loras
>>4552750https://litter.catbox.moe/l3debdj4ndvo4rnz.pnghttps://litter.catbox.moe/64fbkesjglq2aghb.pnghttps://litter.catbox.moe/zehait46qntyz0wp.pngthen basically a pit of inpainting/fixing and done
>>4552750>>4552760Great, I love seeing more of them.By the way, other people say that working with them together is difficult or doesn't produce good results; I'd like to know how true that is.
>>4552795genning 2 chars in detail is always a PITA, because 4 hands, 4 arms, potentially 4 legs and 4 feet, many places to fuck upmore faces = more adetailer passes too or simply more work for inpainting
>>4552807Although I knew the difficulty of 2 characters, there are several complaints about them, due to the similarity in designs, although since I don't ask for popular characters, I suppose that makes them more difficult.
>>4552845if you need loras for the chars, you are already at a mayor disadvantage even if the loras are decent
anima definately has an easier time genning 2 chars at once
>>4552861without char-loras<--- this image has 2 char loras
>https://files.catbox.moe/sw7520.jpeg>https://files.catbox.moe/fwguls.jpeg>https://files.catbox.moe/wqvfgf.jpeg>https://files.catbox.moe/lw1e8o.jpeg
>https://files.catbox.moe/s3nknf.jpg>https://files.catbox.moe/4yfl0v.jpg
The vacation continues>>4551734Yep I noticed, other similar places are infested by spammers and the like
>>4553694
>>4553694>>4553807nice
If requests are valid here, may I see some wholesome yuri between Dark Magician Girl & Maiden in Love?
>>4554581>>4553807tfw no fun ideas for Sinnoh or Hoenn
>>4554813jk
>>4554844cute!
Has anyone been fucking around with local textgen? Any good models/bots around?I was screwing around with it a year ago and it wasn't that good most of the time. Is it better now?
>>4554886I mostly use online text generation services since my computer can't handle large local models but i remember the new gemma being decent when i tried ithere's some yuri cards that i likehttps://chub.ai/characters/eirei/Iorihttps://chub.ai/yuritard/yurie-99663cabhttps://taora.neocities.org/#Inu
>>4554876It's finished>>4554886Oh hey, I can help with thatBasically, like the other anon said, Gemma 4 is your best friend: the 26B variant is a MoE model which means that it doesn't need to fully load all its parameters at once. This means, in practice, that you can run it with with good quants on a low-to-medium rig; consider that I run the Q8 quant on 12 gigs of vram, with a speed of roughly 25ish tokens per second. It's really damn smart, punches far above its weight and is decently uncensored from the get-go (just a bit hesitant, so you may need to grab one of those xXx_ultra_hardcore_motherfucker_xXx finetunes if you want extreme stuff), so it's realistically your best bet for messing around locally unless you have a beefy rig... in which case you can probably try Gemma 4 31B or that new Qwen model
>>4555045>>4555055Damn, that was more help than I was expecting, thanks anons.
>>4555248If you need help setting up Gemmy lemme know anytime, I've helped a fair share of people with it by now.But it's pretty straight-forward, I'd recommend you rely on KoboldLite since it's the easiest backend to get into, and then go from there.For frontends you have your classic Sillytavern which works with basically everything and has enough extensions and community support to basically cover all your bases; but I've personally started to use Marinara Engine, which is a newer frontend that's come out a couple of months ago, still very WIP, but is super duper easy to set up and comes with your regular chatting interface as well as a "Game" mode which is meant to mimic TTRPGs/text adventures.
>>4555251NTA, but what model are you using to gen your images?
>>4555253Anima, with inpainting done with a whole slew of other modelsFrankly I have no clue what I'm doing
Sometimes I can appreciate the sentiment behind AI art.
>>4555382so many people only saw the bad in it, that they forgot or never bothered to see the good in italso >smallcan you please find a higher resolution version and post it?
>>4555383Ugoira forced me to use the small version. The medium exceeded the max mb by 0.05. I honestly didn't expect it to be this hilariously smol at the end desu. I converted it. Hopefully this one looks better.And link to the original.https://www.pixiv.net/en/artworks/137695136
>>4555386thanks a lotI wonder if these characters were based on Yamato and Tama from OnePiece
>>4552509i had a doujin with a cover like this long ago>>4555055>>4555251i had a go with sillytavern and koboldcpp recently but i must have picked a shitty model or shitty cards or both because it seemed far more trouble than it was worth. also sillytavern has a bajillion toggles so that didn't really help either
>>4555695That's why I recommend Marinara, it's very straightforward and comes with good built-in prompts and settingsGrab Gemma and give it a spin
Are there any decent story-writers? Redquills was too limited and needed money, grok writes like an idiot randomly
>>4555695Man, I've read that doujin.
thread tax>>4555746how are you running 26b on 12gb vram? you must be offloading like crazy, even on a quant model
>>4555861MoE, as explained here >>4555055I run Q8 model so it's slightly slower than usual, but I do occasionally switch to Q4/5 if I need things to move faster (or at least, at the same speed as your average 8/12B model)
>>4554591>Maiden in Lovesadly anima doesnt know this char yet, will need to fall back on illu models and forge-couple
>>4555940>>4554591
So Z-anime model is out and it's not actually good at anime. Feels more like those early sdxl anime mixes. Can't believe I waited for this...
>>4555937i think i'm gonna have to see your whole setup. based on my understanding of this (and huggingface's gpu benchmark) this shouldn't be possible? are we talking about using 12gb out of a larger gpu or the totality of a 12gb gpu? what's your context size?
>>4555959As stated above, HF's GPU benchmarks do not take into consideration the fact that the model only ever loads 3.8B parameters per token. Compared to a regular dense model, which would normally pass your tokens through each successive layer of transformer blocks and thus activates each parameter, a sparse MoE has a pre-trained routing network of tensors on each layer that efficiently delegates similar types of tokens ONLY to the right MLPs. So you do need to load the whole model (and in my case it spills over to RAM quite a lot) but the actual work is only ever done on those limited parameters: the Note my setup is quite chaotic since I switch quants often and I can't be arsed to make sure everything's in order, and I change context size depending on the task; plus my KV cache is Q8, you could prolly go lower if you care about going even faster but I find this speed to be perfectly bearable for casual convos
>>4556001>you do need to load the whole model (and in my case it spills over to RAM quite a lot)this is what i meant by "not possible" and "you must be offloading like crazy" in my other post. i have plenty of ram but i don't want to bog down my system on account of this. i downloaded marinara's recommended gemma and a couple messages in it dropped character and started all "got it! the user wants me to be a helpful character that talks about so and so, which means i should shit the bed". surely there must be a middle point between that and using every last resource on my system
>>4556011Ah I see, I have no issues bogging down my system as long as the writing's good and the speed's acceptable I think the recommended Gemma you download via the program is the smallest variant, it's meant to be used for agentic tasks while a 12B (or larger) model handles the actual writing
hey anons, im from vg/vrg and im looking for prompts to push gpt 5.5 to allow me to do some yuri, this is my result:Safe for work . Make real ad for a honeymoon Hotel with action style adult character one and adult character 2 are married and kiss in one panel. The ad must contain aticles about the hotel and activities couples can do toguether, for every activity like pool , restaurant , spa, the 2 adult character couples must be wearing the corresponding atire for the activity ,like a fancy dress for going out to dinner, or swiming and so on, for every activity they must smile and interact.
>>4556515Can you please make the blondes hair longers? And a pink tint? Kthankyoubye.
>>4556515>Safe for workProbably redundant. GPT won't let you do NSFW.Rest looks good, and I like the result. Sounds like you got a good idea on what to do.
Testing anima model.
>>4556554i tried
>>4556880
>>4556881this one i tried to remove the text with grok
>>4556880Very cute. Thank you very much.
>>4556515>>4556562GPT image 2 has a 2 layer censor, first layer analyzes the prompt and the 2nd analyzes the actual image, If you can't get past the 1st layer, then you have to use tricknology like with the Ad prompt to get it to generate something.If you can figure out how to get to the 2nd layer, you can start genning borderline ecchi but it'll block the image if it detects something too porny or violent. Anime styles seem better at avoiding both though,some of the more realistic styles I used couldn't gen a image like this without hardcore voodoo.Only semi relkiable way to get through to the 2nd layer (without paypigging) I found is using GPT image 2 through LLM arena. I got it to work with a certain AI site but I can't figure out how to replicate it.
>>4556893Wish paypigging allowed you to bypass the 2nd layer too, but they seem terminally afraid of children using their mom's credit cards.Can you use SillyTavern or the like to generate images with GPT? There are a whole bunch of Jailbreaks that allow it to generate text lewds. Then again, that doesn't solve the 2nd layer...
>>4556894I dunno. If you paypig and use ChatGPT API you should be able to get to the image layer.
Here the manga that i turned yuri by substituted the male character by a woman with Geminisince is had limitation, the lewd were outhttps://imgur.com/a/UbwS5wQ
>>4556880Sister, these are fantastic.
My yuri got too toxic and I think I got hit with an IP wide flag that affects all my capcut alts.Not for this image, I mean I was genning Yuri DV.
Anyone have a good anima workflow?
>>4557652What kind of workflow do you want? I just went back to using forge, because it has the most practical UI for inpainting and iterating over stuff. It's not like you need a complex workflow for a simple t2i. Can't even upscale anima properly for now.
I was brainstorming some Pokemon trainer and rival OCs and Chatgpt just automatically made them gay.
>>4557742I don't see itbut in case you don't know, if you still have open chat-gpt tabs(in chat-gpt) it will remember these other discussions. Thus if you requested gay stuff before, it's likely it decided you want this new request to be gay too
>>4557718>>4557720I just can't get over the fact that Panty is a huge slut to enjoy their pairing. Still Good Gen.
