>>2753013 hit the image limitPlease keep loli on >>>/b/ and western styles on >>>/aco/asdg> Where do I start?https://rentry.org/sd-mashup> What's with catbox?4chan strips metadata containing prompt info. When someone asks for catbox, they want you to upload the original image with preserved metadata. Prompt information can be loaded from the PNG Info tab of the WebUI.Catbox is free and doesn't require registration.https://catbox.moe/Also consider litterbox for temporary storage.https://litterbox.catbox.moe/> Models and Mixes>>>/t/1190826https://rentry.org/sdmodelshttps://rentry.org/hdgrecipeshttps://huggingface.co/WarriorMama777/OrangeMixs/tree/main/Modelshttps://huggingface.co/andite/anything-v4.0> LoRAshttps://gitgud.io/gayshit/makesomefuckingporn#lora-listhttps://rentry.org/hdglorarepohttps://rentry.org/trashcollects> Recommended ExtensionsThese can be found in the Extensions tab if you're using Automatic1111 WebUI.sd-webui-additional-networksultimate-upscale-for-automatic1111sd-dynamic-promptsa1111-sd-webui-tagcomplete> Recommended models for /e/ contentAnything-V3.0, NovelAI, BloodOrangeMix, AbyssOrangeMix2_hard, nutmegmix, g-anonmixAnything-V4.0 (not actually the successor to Anything-V3.0, but still decent)> ResourcesBad prompt fix embed: https://huggingface.co/datasets/Nerfgun3/bad_promptCatbox 4chanx extension: https://gist.github.com/catboxanon/Big index: https://pharmapsychotic.com/tools.htmlTags: https://danbooru.donmai.us/wiki_pages/tag_groupsHighres fix calculator: https://preyx.github.io/sd-scale-calc/LoRa Training guide: https://rentry.org/lora_train/g/ anon's advanced merging tutorial: https://rentry.org/BlockMergeExplainedModel similarity calculator: https://huggingface.co/JosephusCheung/ASimilarityCalculatior> Related>>>/aco/asdg>>>/d/ddg>>>/g/sdg>>>/h/hdg>>>/trash/sdg>>>/vg/aids>>>/vt/vtai
>>2757488It depends on what you've been doingDid you use highres fix? Either change your upscaler to something like R-Ersgan 4x+ Anime6B or set your denoise to 0.5-0.6 if you haven'tDid you use Ultimate SD Upscaler extension? Hit me with your setting and we can work from thereDid you not use either of those and just set your resolution that high from the beginning? Well, stop doing that and pick one of the above method
>>2757497How'd you get the more cartoonish styles in the top three on the left vs. the more realistic in the further left?
>>2756772How'd you get the water drops on the breasts and body to work out so well? img2img? Inpainting?
>>2757506They're all the same prompt with some settings variations in an XYZ plot. Top left is Anything-V3.0, bottom left is AbyssOrangeMix2-hard.The positive prompt is just "princess zelda, nsfw".
Anyone get inverted nipples to work?
>>2757507several tags. shiny_skin, wet, sweat, cum_on_chest, soap_bubbles, water
Can someone anime this up
>>2757542I want to humour you anon, but without the shiny new toy (ControlNet), this is all I can do
Happy Valentines /e/!I made a love letter lora for valentines, it's a little gacha because hands are involved but I wanted to get it out for the holidayhttps://mega.nz/folder/OgIUSJJD#zo37pMVjdqx-H9b9U3eVEwI also made a bear panties lora I've never posted here you guys might be interested inhttps://mega.nz/folder/mk5lzRbQ#tdCmw4zwRd0zjWkobNF_iw
>>2757560nice. I'll be sure to mess with these later.
>>2757552Wow thanks these are beautiful, or
>>2755412I want to marry her
If I want to train a lora, how do I grab all the images that match a specific booru tag? Are there tools that work for this or should I just consign myself to a lot of right click -> save?Or are there tools that can train a lora based on just inputting booru tags and letting APIs do the rest?
https://files.catbox.moe/771nr0.png
>>2757560I can't get this to work with the character loras I tried (Lum, Mirko, Shantae). Seems to have a strong propensity for making hair pink. Also I noticed a few where the girls were facing the wrong way. I suggest adding "facing away" as a negative prompt.Not that it's bad or anything. Pretty neat for what it can do. Pic related is mixed with a shortstack lora.
>>2757745works alright with the mirko and shantae embeds for me, I don't have Lum so I can't test her. I don't know what checkpoints your using but it works best for me with 7th layer / AoM2You might also have a bad prompt embedding or a generic positive/negative prompt word that's messing it up, you can try increasing the emphasis of love letter also.playing with it more it's probably a tad undercooked, I made it in the span of ~4 hours and didn't properly test it because I wanted to get it out today and was unsure If I'd have the time/drive to continue finetuning it later in the day, I can upload the dataset somewhere if someone wants to take a crack at improving it
>>2757763I tested a few models that were Anything-V3.0 or based on it, didn't try the orange mixes.The bear panties one is good too.
>>2757770glad you like it! Wish I could have figured out how to make the faces more consistent but I think I'm fighting more of stable diffusion as a whole rather then my training settings for that one
>>2757497may i know which model is the second one below from the right (the only one where zelda has clothes) it seems to me the best as an 'anime' style which i prefer to illustration.
>>2757775BloodOrangeMix, my friend.
Why is this thread so slow while I can't keep up with the one at /h/?
>>2757800clearly youre not posting enough. kidding aside, i dont know about others but i dont like to massproduce stuff.
>>2757800/e/ has been a slow board for years now. Let's not overwhelm it with mass produced AI art.
>>2757808>more than 12000 images in output folderI have a problem.
Can someone anime edit this with high heels
>>2757813Fair enough, at least last thread died not due to inactivity but because of image limitAnd I wasn't talking about flooding this thread with image, but it'd be nice to have more discourse about usage. We got questions asked far less frequent here compared to the /h/ thread, for example
>>2757519There's this deep fried lora, I haven't tested its efficacy thohttps://civitai.com/models/8145/inverted-nipples-lora
Requesting more Yuzuriha Inori from Guilty Crown.
>>2757815Mind I ask what models, LoRAs and prompts are you using to get that art style? Really like it.
>>2757821/h/ is known for being the pioneers of coom technology, so everyone flocks there. Plus you can post /e/ in /h/ and the worst thing that'll happen is an anon might link you somewhere else. I haven't seen anyone get banned or even warned for posting softcore there, not many people want to have multiple tabs open for a single subject also
>>2757828>pioneers of coom technologyi like the sound of that. suddenly i feel so respectable like schoolar or gentleman
>>2757828I like the prehensile toes. Drive the mouse with your left foot and keep hands free for more important business.
>>2757828>>2757821Yeah, /e/ is a pretty slow board and /sdg/ is already one of the most active threads. Even so, this is a pretty small thread in the scheme of things with only a handful of regular anons. As far as the technical aspects, personally I'm not too into the training side, it's all I can manage to figure out how to use this stuff. I do like to experiment with what's out there so you could consider me a tester or explorer.Perhaps /e/ just overlaps too much with /h/. At least /d/ and /trash/ have reason to specialize their niches, but for the most part what works for /h/ works for /e/.For now I think /e/sdg works as a showcase for the best things the tech can currently do as well as advice for how to get those results. We're probably not gonna ever be known for being at the bleeding edge of the technology.
>>2757843Also, /e/ has a different activity cycle than the fast boards. This thread will experience a flurry of activity until reaching the autosage image limit. After which it will slowly move down to the bottom of the board and thus out of sight, out of mind. After leaving page 1, it's down to the regulars that actively seek out the thread.
>>2757827https://files.catbox.moe/s69n3z.pngJun lora .3Kairunoburogu lora .6Ootomo Takuji lora 1 7th v3_a model.
Has anyone figured out how to make holographic subjects? I gave it a go but closest I could get still had a lot of visible flesh/skin tones, and I couldn't get there without having to use "slime girl".
>>2757780thank you very much anon, much better than anything 3.0 in my opinion do you know if the nai on /t/ is an updated version of the one on voldy? i still prefer it for inpaint, and would like to have it updated
Messing around with the controlnet extension, It's pretty cool but also really janky. openpose and canny seem to be the best but canny basically rips it 1:1 and openpose only gets the pose right on like 20% of the images I tried
I've followed the guide to install Loras but they don't appear in the dropdown menus in the txt2img UI added by the plugin. Any idea what could be? LORAs are from Civitai
>>2757879The nai on /t/ is the same in the voldy guide. Someone can correct me if they know otherwise, but as far as I know there was only one nai leak.>>2757903Are you trying to use the loras with the native UI or with the additional networks extension?
>>2757903The Lora plugin uses a different directory than A1111 webui's Lora directory. If you continue using the Webui's you can click on the center button under Generate to get a list view of all with cover images (if set)If you're using linux you can also soft link the directories together, you can -probably- do the same on windows, somehow.
Are requests allowed in the threads?
here's a cute outfit to trymidriff, sleeveless turtleneck crop top, micro_shorts, side_tie_panties peek, heart-shaped zipper_pull_tab, o-ring, chain, cleavage_cutout, studded belt, black_thighhighs, fingerless_gloves, armbandadd azur_lane it helps
>>2757947
>>2757946OP here. Requests aren't generally encouraged. How to questions are welcome, but they should be about learning and not just "do this for me".Nothing wrong with anons fulfilling the occasional request if they feel like it, but I don't want this thread to become a request thread.
>>2757970>why would you be requesting what you could just make yourself?Downloading all the shit is compwicated
>>2757970maybe he does not have the hardware for it? pretty sure i saw an online version of novelai somewhere, or that chinese site. anyone has a link?
>>2757944>>2757945Thanks for the help, all fixed now.
>>2757999This is the one I use before I daringly install locally in my 3050 4GB VRAM laptopcolab.research.google.com/github/acheong08/Diffusion-ColabUI/blob/main/Diffusion_WebUI.ipynb
>>2757952Fair enough, I was the one responding to the captain america-looking girl request, but yeah I wouldn't like it either if this thread is flooded with request, especially difficult onesWouldn't mind generating some simple concepts tho, sometimes I just don't know what else I should generate
>>2758011Perhaps a good middle ground is to announce that you're looking for inspiration or taking requests. Otherwise, requests out of the blue are gonna be ignored because no one's in the mood.
>>2758016You're right (possbily) OP. With that being said, I'm volunteering myself for a simple request that doesn't call for obscure poses or characters. Since I don't actually use this thing to coom, I lost motivation to experiment sometimes
>>2757828JFC I miss Salama
>>2757946I've thought several times about opening a thread dedicated to requests, I've never done it, because if someone has sd, like this anon pointed out if they do it themselves, and it's pretty hated on the board (outside of here) if there are others who want to make requests I can think about it, as long as the concepts are simple, and mostly original character over copyright character
>>2758035>this anon pointed out >>2757970
>>2758011If you don't mind, what tags should I try for this composition?
>>2757821>>2757505Catbox? Would like to try and make something similar to these
>>2758041lying, spread legs, armpits/hands behind head, bedI used these, if it doesn't work just strengthen one or some of it
I've made so many girls with large breasts in tank tops with denim shorts, yet I never made a Misty.
>>2758067I was actually experimenting when I made those, not to mention inpainting, so the catbox wouldn't really reflect what I usually do to make things similar to theseWhat would you like to see?https://files.catbox.moe/3b0ghr.png
>>2758095Can't make, won't make or just not interested in making Misty?
Cute Bnnuy
>>2758099Just not interested. I tried one early on but it can't decide between anime and game.
>>2758094What the fuck is this asshole doing in that armpit
>>2758097The edges are pretty crisp/soft (where they need to be), and the top being pulled taut looks pretty good instead of just being a solid block of boob
>>2758114Holy shiet I've never noticed that>>2758122Then that's probably mostly due to the combination of model (Kotos abyss proto) and sampler (DPM++ 2M Karras)https://civitai.com/models/5245/kotos-abyssproto
>>2758124>samplerNow that it's been a while, is there a decent guide to what samplers do what best instead of just>uhh, here's a bunch of pictures bro, figure it out yourselflike some are better with lighting, some don't fuck up hands, some make nice backgrounds, etc. etc.?
>>2758180Euler a is the king of speed and should be your goto for seed hunting All of the DPM++ samplers are good, DPM++ 2M is probably the best upscaler if you crank the steps up high (50+) DDIM is good at blending during inpaintingadaptive sucksThere's no be all end all for samplers because of the nature of the seed gacha, you can generate x-plots of the samplers to see what one looks best on the seed you're on
cant stop wont stop
>>2757819Bump
>>2758182>go Euler a for seed huntingBut if I switch to DPM++ later after finding a good seed with Euler a, wouldn't that change the image composition as well?
>>2758205Yeah it will, you don't need to find the perfect seed with Euler a just one that fits the image in your head enough to warrant investing time running plots and tweaking settings on. With a good prompt the composition shouldn't drastically change. It's debatable if seed hunting is even worth it.welcome to gacha, enjoy your hell
Anyone tried the new AOM3 yet? It was released about an hour ago.
>>2758211I've downloaded the base model, but I haven't messed around with it yetThough I suspect i wouldn't see a huge jump from AOM2
>>2758211I've always liked the quality of AOM2 but not necessarily the style. Interesting that the A1 variant is specialized for a flat style, I'm gonna give it a go.
>>2757882I just installed the controlnet extension. Does anyone know what models I'm supposed to put in the model folder? There are the ones linked on github, and the name of those implies they're made with or optimal for SD 1.5. Are there other models out there? Will any model work or is it something specifically for controlnet?
>>2758233Their specific models, only other models I'm aware of are the pruned versionshttps://huggingface.co/lllyasviel/ControlNet/tree/main/modelsand the AOM2 mixeshttps://civitai.com/models/9557/abyssorangemix2-controlnets
>>2758238oops first link was meant to behttps://huggingface.co/Hetaneko/Controlnet-models/tree/main/controlnet_safetensors
>>2758238Thanks anon. The dude that makes these orange mixes is very interesting.Also, I'm giving AOM3 A1 a try and it makes fluffy Mirko pubes, which makes me happy.
I posted in /g/ but maybe this thread knows - Is there a way to get a garment to have specific characteristics in a prompt without fucking up other parts of the outfit?Daisy Dukes, for example - they're cutoff denim shorts, usually the length of hotpants. Is there some way to tell SD "these are denim shorts, they are cutoffs with frayed hems, with the length of hot pants"? Kinda like nesting properties, an AND statement, something like that.
>>2758250Most consistent way is probably to train a lora/embed of them Could also find an image of someone wearing it and run it through controlnet with canny/depth/scribble, It'd be faster then training but would lose almost all flexibility, you'd essentially just be tracing the image.If you want to get them through pure prompting your best bet would be to increase the emphasis of the tag(s) and pray to the gacha. Daisy Dukes for example is tagged as denim shorts in danbooru, so most nai based mixes (almost all the 2d ones) are going to associate it with any type of denim shorts without further training
>>2758252The denim shorts tag is the kicker - if I could just do "denim shorts, frayed cuffs, hem at upper thigh" it'd be fine. If I just do denim shorts, how does the NAI model pull a specific type? Does it randomly select something based on the seed, e.g. "denim shorts" generates near capri length shorts on seed 12345 and basically hot pants on seed 67890? Or is it truly random based on whatever the model thinks is best?
>>2758254I only have a layman's understanding of how it works, but its concept of "denim shorts" is just an abstraction based on everything fed into it with that tag. It's not like specific seeds will draw out a specific image. It has a concept of denim shorts based on the sum of its input. The limitation lies in the granularity of the tags it was trained with. If it was trained on a lot of shorts and only a few with frayed edges, it will have an overwhelming propensity to make them without frayed edges.This is why a lora is so helpful. You are refining that training and feeding a better dataset in, so now it has a better concept of "daisy dukes".On an unrelated note, here's a Lum from AOM3 A1.
>>2758255yeah that's basically how I understand it as wellcute lum
We call things like this AI because the closest thing we have to describe what it does is the word "learning".Think about how a child naturally categorizes things. That's a dog, no that's not a dog that's a cat. What's the difference between a dog and a cat? As the kid gets more input, he will be able to distinguish between a dog and a cat, then between breeds of dogs. In your mind's eye, when you think of a dog, you likely don't think of a specific picture of a dog, but you have a concept of a dog synthesized from all the dogs you've ever seen.In the same way, these models are only as good as their input.
the lately added latent upscale option only gives me blurry results. i tried searching around but cannot find anything on it. anyone has an idea why that is?
the pillar want give you some love
>>2758303i dont know what i did but it fixed itself. maybe some memory issue.new aom3 is looking sharp btw.
The infinite power of controlnet....
>>2758383very nice
>>2758401valentine's day.. chocolate gifts.. melting. Really nice anon, I liked this one.
>>2758303>>2758326Latent has always blurred my images. Your best bet is to grab an 'anime' intended upscaler and then editing the ui-config.json file so your new upscaler is the default
Chanced myself into a 4090, are there any good "throw power at the wall" type settings I should know about? I have no idea how to get the most of this thing
Skirt lift is the best Lora.
>>2758464Is Not!Priestess anon still active?
Sav'aaq
I cancelled my NovelAI subscription months ago because it couldn't figure out how to make slightly more obscure characters like Makoto Nanaya. However, I've lately noticed more AI art of other characters which I failed to make back in October, so it seems it's getting better at specifics. Has AI progressed since then to the point where it can make Makoto Nanaya?
>>2758464Use --no-half --no-half-vae when launching. Other than that, you're just capable of higher resolutions before running out of memory. In the world of AI, it's not cores or clock speed, VRAM is king.>>2758565A good enough lora is capable of making any character using a handful of images as a starting point.
>>2758326https://files.catbox.moe/zu7ppp.png
how do i register for civitai? when i try to sign up it says use the api but checking the jsons for everything without having pics to see what im looking for is bullshit
This controlnet thing has potential, though I'm still getting the hang of it.
Has anyone here trained a LoRA to mimic a particular style? I've just managed to follow directions enough to produce a working output, so tuning it to work *well* seems like the next step. For this first attempt, I just naively dumped and tagged 20 images as the training set. Are there any organizational or tagging best-practices when trying to reproduce a style?
>>2758783https://rentry.org/2chAI_LoRA_Dreambooth_guide_englishhttps://rentry.org/lora_train#using-lorasThese guides helped me a lot, long-winded though they are
>>2758784I'll spend tomorrow giving them a read, thanks!
More Not!Priestess please.
>>2758581Good to know, I'm keeping an eye out for those zany A series cards if I can get one at a good price >>2758497Sorry anon, I'm just a passing stranger
>>2758492Where the heck is it? I looked through all lora links in OP
>>2758821https://civitai.com/models/8631/
wow it actually got the spelling right
>>2758950>>2758951good stuff, but you gotta inpaint those hands
>>2758952...how to impaint hands? i know how to move it to impaint and sketch over the hand but, the whole image comes out blurry? And how to get a better hand pls.
my Note lora is coming along nicely
>>2759021Nice, looks pretty on-model. What model and are any loras involved?
>>2759026Midnight mixer melt modelFor LoRAs there's a Jessica lora on civitai, and I also used a Toriyama artstyle lora
I need your muscle girl prompts
>>2759123
>>>/vt/43450751Nice Idea.
>>2759172Welcome to /e/. What advances in /vt/ technology have happened lately?
>>2759175Not the anon that made the "Nude Lulu in Space" but the one the post is responding to. /vt/ is doing 'things' with theis chuubas. Inexplicable, housewifery things.
>>2759180Her eyes make her look like she's been bonked on the head. Surely you have not been bonking your waifus on the head.
reverse bunnysuit + raita style
>>2759156more pls
>>2759236God fucking dammit. I'm going to have to learn this shit.
Any inuyasha stuff
>>2759253Bitches be crazy
>>2759254I don't know how this is related to /e/ but ok.
>>2758464How'd you get the one arm up?
Any advice on getting tanlines to work without getting cubist stripes?
>>2759252right? these sluts are crazy good
>>2759392I've used the arm_up tag to good effect (not to be confused with arms_up)
>>2759401No idea. I can't even get a lineless tan without using bullshit like "ryukyu sun touched", it always ridicules my skill like this as soon as I have "tan" in the prompt.
Beating poison at strip poker!>>2759401Prompt for dark skin (there's a bleached lora that can help with this if you're having trouble getting chocolate girls) and sketch inpaint the tanlines, there's to many variations so the AI has an aneursym when it has to draw them blind
>>2759423Someone send this to Jean-Paul Gaultier and tell him "I dare you to make a swimsuit that does this tanline"Meanwhile, subject change - as cool as it is to go from an image to modern art, what am I doing wrong with upscaling to get this original...
>>2759428... to turn into this result?All I'm doing is checking off "hires fix" and leaving the defaults, I haven't actually been satisfied with any of my generations to try upscaling yet, so it's still new to me.
>>2759430your denoising strength is way too high. It's usually good somewhere around 0.3-0.5
>>2759441Gotcha, thanks. I kinda like what it's doing, a very surreal sexy. Someone merged Dali at some point. Before:
>>2759446After
>>2759447the thumpnail looks like she exploded due to a real fat creampie
>>2759369Watching this while drunk or high must be trippy as fuck>>2759430If you're not satisfied with your results, seems like you need a new model my guy, what's your current model? Base NAI?
>>2759489I was satisfied with the results before the upscale. This one's using Anything 4.5, I didn't want to futz too much since the background and color scheme, tone, etc. were perfect. I'll probably take a fresh shot at it if I can stop myself from cooming while generating, it's hard to focus on improving and tweaking when your own art makes you want to fap.
>>2759243Here you go
>>2759505
>>2759506
>>2759507
>>2759447jeanie
anyone has a tip on how to generate nice buttholes? ive tried for 3 days to get one on a picture that i like, but all look kinda mangled, gaped or somewhat out of place. i already looked for a lora but could not find one.
>>2759560Have you tried the AbyssOrange hard model? Might be worth a shot if not.
>>2759621Do you know what board you're on? How is posting only male subjects on topic?
When I try to reproduce the exact image, with the same catbox properties, on the same model, it turns out to be different. Why is that? If everything is set up the same, it should produce the same image right?
>>2759712Same seed and everything? I've heard of people not being able to reproduce the same image if there's a difference between AMD and Nvidia. I'm not sure what else would do it.
>>2759712>>2759722At least locally, I've seen that images I post-process in img2img (be it inpainting or whatever) end up saved with whatever the last seed used was, clobbering the original. Could be that?
>>2759742If the source inpainted stuffs, their meta data would change accordinglyYou can only get a 1:1 if it's pre-inpainted version
Has anyone tried or had success colorizing manga pages? I'm training a LoRA on a particular artist and just realized I may be able to do this to expand my training set.
>>2759772i sometimes do edits in the photoshop thread here on /e/. i can do almost all kind of edits using sd, only thing where i had to give up was coloring. you can do it but only while greatly transforming the picture. so much that its an completly different picture. that beeing said i know there are models specificly for inpainting, maybe there is one for coloring too but i could not find one.
The first cup of coffee in the morning is the nest one.
For me, it's princesses in latex suits
>>2759871Are you using the bad_prompt embedding? If not, check it out so you don't end up with handnipples or whatever's going on with the coffeehttps://huggingface.co/datasets/Nerfgun3/bad_prompt
I finally figured out inpainting and got enough time and focus to work on something I'm kinda proud of. I lost the original prompts through inpainting so catboxing won't do much. For the curious, the detailspositive prompts: masterpiece,best quality,highres,best illustration,4k,1girl,(ritsu tainaka:1.5),(hairband:1.5),clear hazel eyes,warehouse,(topless:1.5),(pert breasts:1.5),(jeans:1.7),grin,smile,(nipples:1.6),(one knee:1.6),crossed arms,short hair,(tsurime:1.5),(forehead:1.5),(converse:1.5)negative prompts: cropped,simple background,missing hand,amputee,extra hands,cripple,mutated limbs,menu,ui,ux,user interface,mismatched eyes,angry,malformed breasts,large breasts,huge breasts,extra limbs,hair bobbles,hair ribbon,ribbon,capri pants,(bad_prompt_version2:0.8),pussy,unzippedModel: anything v4.5Sampler: DPM++2M KarrasSteps: 46Width: 512, height:728CFG: 8.5Seed: 2348579659Had to do a round of inpainting to remove weird folds on the thighs of the jeans and hide a protruding ribcage. Took about a billion tries to get a proper Ritsu hairstyle and hairband combo or at least as close as I could get. I'm open to any constructive criticism that's out there, this thread is a fantastic font of learning
>>2759903That one used EasyNegative rather than bad_prompt. Hadn't spotted the hand-nipple on what was otherwise an OK minimal gen.
>>2759904looks realy nice. in the options you can set sd to generating a txt file with the parameters next to the img in your output folder. you also can include model info in the metafiles. anyone should have that one checked imo.the one info i found most helpful when inpainting was that you can just paint a sketch of what you want in paint or photoshop and then inpaint that. that helps a lot with fingers, eyes or any other edit you want. pic related.
>>2759906NTA but expecting average people to draw hands/fingers on photoshop that's good enough to guide the AI seems like a tall order when many artists are struggling with it
>>2759712I think so. I've put the desired catbox image to PNG info, and I've sent it to text2image.Image I wished to reproduce: https://files.catbox.moe/a4u5ly.pngImage I got:https://files.catbox.moe/e9h6zv.pngYou can see those two have identical properties, except one model name is upper case, but the same hash. Idk if I'm doing something wrong, but I can't seem to pass the sanity tests for any model I download.
>>2759772You aren't alone in your interest.I've tried a little but with no acceptable results. Controlnet seems to be the appropriate tool, I just haven't gotten the hang of it yet. I'm also not sure of the best workflow for it. Will it work on whole pages or do I need to focus on a panel at a time? Who knows. If anyone has tips I'd like to hear them.
>>2759984I'm gonna mess around with it this weekend, I was planning on isolating single panels / subjects in a pre-processing step. I think the secret sauce here is that I've already trained a LoRA on this artist's style, so I might be able to use that bootstrapping to work around whatever transformation issues >>2759814 was getting.
>>2759954no not at all. u dont have to draw em. just a rough sketch in a flat color. the outline is all you need to get decent results. give it a try. i draw hands like a 5 year old and it works.
>>2759966when you fail the test on any model must be you. i forget what it was called but you are supposed to set delta denoising to something like 31442. ill check later what it was. maybe try updating your webui
>>2759966>>2760042https://rentry.org/nai-speedrunhttps://imgur.com/a/DCYJCSX> Eta noise seed delta = 31337this is what i meant.
>>2760036Author of the Mugi post here, my hands secret is quite literally just the bad_prompt embedding. Sometimes it works with version2, sometimes wtihout, but as long as it's in there and I'm not having the hands explicitly do something in the prompts, it's been OK. Not perfect, but OK. >>2759983Impressive detail on the stocking tops, any prompt tricks to induce it? I've had some luck with lace-trimmed legwear but nothing really consistent.
>>2760090nope nothing, just "thighhighs." the model for that one was anythingv4.5
Any one have an inuyasha one
>>2760238
>>2760239
>>2760240
>>2760241
>>2760242
>>2760243
>>2760244
>>2760245
>>2760246
>>2760247
>>2760248
>>2760249
>>2760250
>>2760251
>>2760252hand could be better on that one, should just use img2img and fix
>>2760253
>>2760256
>>2760258
>>2760260
>>2760261boota
>>2760262
>>2760264
>>2760266
>>2760267
>>2760268
>>2760269
>>2760271
>>2760272
>>2760273
>>2760274
>>2760275
>>2760276
>>2760277
>>2760278
>>2760280
>>2760281
>>2760282
>>2760283
>>2760284
>>2760285
>>2760286
>>2758866>kimono>pantieswat
>>2760329While it's traditional most girls don't go commando in kimono or yukata in modern age.
>>2760331Although it should be encouraged of course. By good example and such.
>>2760266>>2760276>>2760277>>2760281Can I get a Catbox?
>>2760355https://files.catbox.moe/1kbvmt.pnghttps://files.catbox.moe/9x381l.pnghttps://files.catbox.moe/ctrxfm.pnghttps://files.catbox.moe/a1j85l.png
>>2759814>>2759984>>2759991I would like to apologize for my hubris, it is insanely difficult to preserve the linework no matter what I try.
>>2760499At least you tried.
Asked this in /g/ and they pointed me this way, I've set my AI up following the voldy guide on sdg. I'm not the most computer literate person, this stuff seems pretty high tech. So far so good, having some trouble with img2img but I'll get the hang of it.What's next for me to really dive into waifu cultivation?What are good anime models where I can easily mimic or synthesis mangaka artstyles?What are hypernetworks and are they hard to learn?What is LoRA and is she hard to learn?And HOW?
>>2760545As long as the model doesn't have a heavily photorealistic bias, you will get anime girls by default. Pic rel is a test prompt for a batch of different models.Hypernetworks are old-tech, forget about them.LoRAs are the new tech for model extensions. To use them you just drop them into the appropriate folder, refresh the UI, and then add them to the prompt. LoRAs add things like different art-styles, specific characters, or items.Building your own is heavily discussed over on /h/
What kind of prompts can you use for multiple people in an image (one nude, and others not)? For group/ENF based scenarios
>>2760551I am not very experienced so take what I say with a grain of salt. From what i've seen any tag you use affects the whole image so you will often end up with everyone nude or everyone clothed. My solution is: lower the resolution to the point you can generate lots. Make some images and see if your prompt produces desired results occasionally. If not, lower the weight on nude tag or the (whatever someone's wearing) tag and rise the opposite one. When you start getting good results just upscale or try resizing seed or start rolling multiple images in bigger resolution with that prompt.You can probably train a lora for this but my toaster can't do that.
>>2760545>What are good anime models where I can easily mimic or synthesis mangaka artstyles?Check the OP for model recommendations.I would start with Anything-V3.0, it's the least fussy and produces good results out of the box. AbyssOrangeMix 3 is also very popular and a good choice.>What are hypernetworks and are they hard to learn?>What is LoRA and is she hard to learn?Hypernetworks and loras are similar in the sense that they modify your output along certain lines like applying an artist style or specialize for a certain character. Hypernetworks have fallen out of fashion in favor of loras because they are more flexible.Using either isn't difficult. You simply download them and put them in the folder they belong. For loras, you put them in the models/Lora folder and refer to them in your prompt with <lora:loraname:1> where 1 is the strength you want to apply, 1 being the default. Increase or decrease it to adjust the influence of the lora.That reminds me that I still haven't made a good xyz plot on the latest popular /e/ models. I guess I'll work on that today.
If anyone has suggestions on what experiments to run, I'm all ears. Always looking to generate illustrative plots.This one I call the Yuki test. Simply prompt for "yuki nagato". Such a simple prompt can illustrate a model's bias for how well it knows the character, how nsfw it leans by default, and other things like how good it is at eyes, what style it leans toward, etc.
hi, is this the right board?
>>2760556Cheers. I’ll get anything-v3.0 and report back eventually with some results and what I’m wondering in order to move forward. I like some of the results I’ve seen from anons using abyssorangemix and that g anons mix. I’ll get them all and have a little play around!Are these models good for creating non nsfw stuff too?
>>2760606If you don't prompt for nsfw stuff, you won't get any. See >>2760547 for various sfw prompts with several nsfw-capable models.
>>2760606>Are these models good for creating non nsfw stuff too?Yes, Some models (the orange mixes for example) are more nsfw oriented than others. Anything 3 will only generate sfw stuff until you start including nsfw tags (nipples, pussy, etc). It wasn't trained as a nsfw model per se, it's just not censored like the flagship sd1.5/sd2.1 models.Pic related is from Anything 3.
>>2760603Style seems on the border of /aco/ but it's fine. The only hard and fast rule is no dicks. Plus /aco/ really only cares about celeb fakes so welcome to /e/.
>>2760603she's a beauty; catbox, pls?
>>2758866catbox please?
>>2760661
>>2760662
>>2760663
>>2760665
>>2760666
>>2760667
>>2760383Thanks!
>>2753658Would you be so kind of sharing the training parameters you used and well the lora as well if you dont mind
>>2760692Sorry about the miscommunication, I merely evaluated a lora that someone else made. However, he shared some notes that should prove helpful.https://mega.nz/folder/s8UXSJoZ#2Beh1O4aroLaRbjx2YuAPg/folder/oo9WhSgL
I love this robot like you wouldn't believe.
>>2760739
>>2760740
>>2760741
>>2760742
>>2760743
>>2760744
All willl bow before the CLAW
>>2760745
squat
>>2759123grainy early 00s concept art of muscle girls is a way of lifethe Shigenori Soejima lora from /g/ makes anything look like persona 3-5 but with low weight it can be wrangled to draw from persona 2https://files.catbox.moe/m0sb3y.png
>>2760715The robot doesn't love you back, at least not enough - her areolae are squares
>>2759906So if I'm trying to remove a visible mistake - e.g. the line coming out of her nipple here - could I just draw a shitty mspaint sketch of a breast then use it as a mask in Inpaint Upload?
Grapefruit seems to have the best lighting imo, although it gravitates toward NSFW generations
>>2760711kek that's my lora, np i just wanted to know what other people were doing to get her right since i kinda just lucked that one out without really giving much thought to the training settings. Nice to know it does work tho. I guess at the end of the day dataset really does 90% of the work.
>>2760856Pretty much; it's there just to guide a new image gen.
>>2760856My workflow for that kind of thing is to just go to the inpaint tab and draw a line over the blemish (with masked content set to "fill"), then generate ~9 images at between .5 to .7 denoising strength.
>>2760911Thank you! Do you keep the same prompts as they were carried over from txt2img or do you just prompt for what it should be?
Does anyone have any advice on what prompts I should use to try and get this hairstyle? I'm using img2imgI've got: parted bangsforeheadcat earsbut not sure what else to use to try and make the tufts
>>2760938I always propagate the original prompts, I'm not sure what you'll get if you change things up, but I suspect the style will start to shift in the newly generated image (you're certainly free to try, though).
>>2760938For inpaint masked, describe what's being replaced (breast of ...), for whole image inpaint, repeat the original prompt.
>>2760946Parted bangs, long bangs, spiked hair maybe?
>>2760946someone already made a lora for au ra, there's got to be one for miquote>>2760962ask for anything spiky and you'll get a girl with chun li bracelets
>>2761122
>>2761123
>>2761124
Are there any models that excel at cyberpunk/synthwave backgrounds and buildings and stuff? Or is it lora-trainin' time? Also, out of curiosity, how long can you guys spend working on generating something without just shifting right into coom mode? I feel like I make no progress once I get something I like that needs refining, it's just fappin' time
>>2761169>Are there any models that excel at cyberpunk/synthwave backgrounds and buildings and stuff?SD1.5/SD2.1 are pretty good at landscapes, cityscapes, and backgrounds in general, and pretty good at cyberpunk style. The nai and anything based models are sorta ok.>Also, out of curiosity, how long can you guys spend working on generating something without just shifting right into coom mode?Most of the time my workflow involves getting a "good enough prompt" then walking away for hours while it generates batches on various models and settings using an xyz prompt. Then I go through and cherrypick the good ones, and do a second round of cherrypicking to upload here. I'm not one to seek a perfect result by continually refining the prompt and honestly I try to tax my GPU to use it as a space heater.
>>2761173Do you stick to one seed on your batches? Also, when you're using XYZ prompts, is it just sampler/steps/CFG as your axes or some other properties? I'd love to be able to have my machine crunch stuff out while I'm not using it, that's a good idea you've got going.
>>2761173Also, dumb question - if I want to merge SD2.1 with Anything 4.5, but I only really care about the landscapes, cyberpunk, cityscapes, etc., is there any way to do so? Or is my best bet just to merge them at 0.5 and hope for the best? Optimally I'd love to be able to build checkpoints that deal with just a specific thing - e.g. a geographical area, view, etc., and make 'em waifu photoshoot background fodder
I realize this isn't ecchi, but the /g/ is consumed by tribal warfare so I hope you guys can help me out.Any advice on what upscaler is best for cleaning up/restoring a low-quality game footage screenshot like this?I've tried several different upscalers targeted at cleaning up jpg corruption from the upscaler wiki, but I've had no luck getting anything decent so far.
>>2761134Really impressive gen, anon. Hair and face especially are just fantastic. What model are you using?
>>2761203I'm no expert on merging, but from what I understand, SD2.1 is different enough from Anything type models to be a headache especially when trying a method like weighted sum. Some kind anon has written a block merging tutorial that you might want to look into:https://rentry.org/BlockMergeExplained>>2761200>Do you stick to one seed on your batches?No. -1 and I check the option to keep -1 on everything generated by the XYZ>is it just sampler/steps/CFG as your axesI usually use different models as my Z axis and sampler, steps, or CFG on X and Y.For example I'll do something likebatch: 6X (sampler): euler a, dpm++ sde karrasY (CFG): 6, 8, 10Z (Checkpoint name): anything-v3.0, anything-v4.5, AbyssOrangeMix3, AbyssOrangeMix3-A1, g-anonmixAnd that will crunch for an hour or two. I also have live previews enabled so that if I check up on it I can easily see if it's gone off in the weeds and is making junk.
And yeah I know the robot girls aren't exactly the best examples. They're just the most on-topic stuff I had for the purposes of the conversation.
>>2761205Thanks, it's "7th Anime v3_A", a subsidiary of 7th LayerEasily found in hugging face, just search 7th Layer Anime
>>2761313Do you know what the difference is between A, B and C?
>>2761314I've been scratching my head on that as wellI think it's mostly on the lining, A is more flat, while the other two are more... 3D like? Look at the loli example in the main page, I use that as reference
Wow this one has made the best Alitas so far. Nothing else has been able to get the lips. (7th Anime C by the way)
Wanted to make an ENF image, instead got horny elf.
I'm working on my first character Lora and currently setting up the dataset of images. How thoroughly should I be pruning the tags auto generated by the WD tagger? I've gotten rid of the character specific tags like black_hair and so on, but how much beyond this should I be concerned about? Some of the images ended up with like 50 tags with a lot of them being somewhat generic ones like lying_down, upper_body, looking_at_viewer et cetera. Should I keep these or will an excess amount of tags somehow negatively affect the training process?
>>2761087I love the look of these. Would you mind posting catbox? What model?
>>2761550Anything 4.5 with these loras:https://civitai.com/models/13721/bluethebone-esthetic-80s-stylehttps://civitai.com/models/10565/inverted-nipple-boobsHere's the images:https://files.catbox.moe/9rsi4h.pnghttps://files.catbox.moe/2v6pq7.pnghttps://files.catbox.moe/fmd9nt.pnghttps://files.catbox.moe/xfgwhb.pnghttps://files.catbox.moe/zf3shb.pnghttps://files.catbox.moe/kln7y1.pnghttps://files.catbox.moe/pgcdpc.pnghttps://files.catbox.moe/ci1cq4.pnghttps://files.catbox.moe/rfansj.pnghttps://files.catbox.moe/y9nd7c.pnghttps://files.catbox.moe/199q1b.png
>>2761411what prompt or tag that make them cover their initimate?
>>2761580Don´t really remember the details, but I basically described an ENF scene as a prompt, something like: "Embarrassed and svelte black-haired woman standing naked on a field trying to cover up."Mind you, this was the best picture of the batch, the other results were a bit like this, so... yeah.
>>2761607Also tried the same prompt in a forest if you are interested in the results. Don´t know where the towel came from, but the AI really loves naked elves apparently.
>>2761726Catbox please?
>>2761729https://files.catbox.moe/qlac5a.txtsorry I'm silly and don't know how to embed it into the image or how that works
>>2759760>>2759810catbox for these?
>>2761735It's actually already embedded in the image. 4chan just scrubs the data when it gets uploaded. If you drag an image into the stable diffusion prompt and click the blue arrow on the right it'll propagate the fields with those settings.
>>2761411>>2761607>>2761608You're missing the VAE, anon
>>2761587wats prompt for this boobs form?
>>2761580For boobs handbra sorta works for me, like quater to half the time when no other prompts hinder it. For pussy didn't find a good one, try pussy in negative, hands hiding pussy in positive, increase its weight and roll for the best.
>>2761735thank you sir where can I get the akira lora?
I made a cammy, I'm loving this stuff
>>2761853I made it, here you gohttps://files.catbox.moe/v11kre.safetensorsI just trained it on whatever model I was using at the time but it seems to work well enough across different models
Is there a good way to create a LORA of the snake princess outfit from shantae? I`m a brainlet
risky streaking under moonlight, what could go wrong?
>>2761992I think you're missing a vae? Shame because it's a pretty good image otherwise.
OP here. Image limit is coming up. I don't want it to hit the limit overnight so I'll make a new thread shortly.
Gonna round out the thread with a couple posts and end it with a crosspost to the new thread.
>>2762003>>2762003>>2762003New thread.
