Side Hustle EditionPrevious Thread: >>>>8751615>LOCAL UIreForge: https://github.com/Panchovix/stable-diffusion-webui-reForge | Stable commit: 6964cebComfy: https://github.com/comfyanonymous/ComfyUI | https://comfyanonymous.github.io/ComfyUI_examples>NOOBAI-XLhttps://civitai.com/models/833294/noobai-xl-nai-xlhttps://huggingface.co/Laxhar>NOOBAI SHITMIXES102d custom: https://civitai.com/models/1201815?modelVersionId=1491533291h: https://civitai.com/models/1301670/291h>PAIDhttps://novelai.net/image>RESOURCES4chanX Catbox/NAI prompt userscript: https://rentry.org/hdgcbEasy booru tag export userscript: https://github.com/Takenoko3333/Danbooru-Tags-Sort-ExporterTags: https://danbooru.donmai.us/wiki_pages/tag_groups | https://danbooru.donmai.us/related_tag | https://tagexplorer.github.io/#/Inpaint: https://files.catbox.moe/fbzsxb.jpg | https://huggingface.co/Wenaka/NoobAI_XL_Inpainting_ControlNet_FullUpscalers: https://openmodeldb.infoBooru: https://aibooru.onlineControlNet: https://rentry.org/dummycontrolnet (OLD) | https://civitai.com/models/136070 (look at the green links in the description)>TRAININGGuide: https://rentry.org/yahbgtr (WIP)Anon's scripts: https://mega.nz/folder/VxYFhAYb#FQZn8iz_SxWV3x1BBaJGbwTrainers: https://github.com/derrian-distro/LoRA_Easy_Training_Scripts | https://github.com/bmaltais/kohya_ss | https://github.com/Nerogar/OneTrainerOP Template/Logo: https://rentry.org/hggop/edit | https://files.catbox.moe/om5a99.png
Rectal flow + gopro lora :D
>>8757780way too anime, turn the gopro up
>>8757780gib gopro dataset
>>8757791Unfortunately, my gopro lora IS anime, unlike that other gopro lora, so if I turn it up more you'll just get more anime...
>>8757780>Rectal flowAre you prompting on cumfee?
Crazy how one little url sets off your mental illness like this.
>>8757801Yeah, switched long ago due to resonance cascade.
>>8757780b0x please?
>>8757805whatever even happened to thatdid he just take the money and run?
>>8757896No, he stayed at home in his fursuit.
>>8757896He kept training on his 970 and invested in his fursuit.
>>8757923>his 970excuse me, he had a riva tnt cluster.
>not training on a cluster of ouya's
Anyone care for Wan gens around here? Downloaded it today and seems like I can actually gen OK quality stuff in just under 3 minutes with all the optimizations available now.
>>8758090post them and don't listen to static schizos
how on earth are you getting such clear details and fine lines?
ever heard about upscale and inpaint lil bud?
>>8758106I just started to use inpaint and I don't know where to start with upscaling
>>8758108If you are using reforge (I really hope you are), you don't need anything more than hires fix with proper settings
>>8758112I'm using comfyui.
>>8758117Then you have a long road ahead of you
>>87581081. Upscale with a model (animesharp is fine, some ̶r̶e̶t̶a̶r̶d̶ will say there's no difference even if you do it with lanczos, use your own judgement).2. Do an image2image with 30 to 50% denoise (or higher with controlnet).3. THE END. (not really, there's still multidiffusion, CN Tile, choice of samplers and lots of other crap to go even higher resolution, but it's THE END for now)
>>8758137>some ̶r̶e̶t̶a̶r̶d̶ will say there's no difference even if you do it with lanczos, use your own judgementi feel like it's just 1 guy who fucking says this.
it's always the 1 guy
1guy AMA
>>8758158why are you gay?
>WAN copeyou love to see it.
>>8757975getpregnantgetpregnantbox?
>>8758146I say the difference between latent and literally anything else (lanczos included) is far greater than the differences between it and all the "real" upscalers.
Requesting catbox for this gen if the original anon is still lurking
>>8758186well that's true, partly because latent basically requires high denoise to functionthe difference between pixel space upscalers is still pretty big
>>8758190I don't think it's from gopro anon (you wouldn't get shit from him if it is anyway).My first guess would be this lora: https://mega.nz/folder/njQmnDQZ#EoKYxBHc29wIMKIWO8Zl1A/folder/HyJSlTyY
>>8758159Explain without sounding mad how trap (formerly otoko_no_ko) is gay, chat.
hehehe YOINK! :3
Im using WAI nsfw illustrious sdxl, is there any way to fix the low eye detail it tends to output?
>>8758524I use adetailer for exactly that but some anon on /g/ said I shouldn't because it creates a halo or something? All I know is it's the only way to improve the image and some dude says I suck.
Cute newfriends
>>8758531Could you point me to that adetailer? I'm just a newfag in here and a simple google search doesn't show up to much
is there a quick way to get a list of something like hololive girls or pokemon but include their default outfits? I have some wildcards setup but they just have names + some features wondering if there is a quick and painless way to get additional details in there.
>>8758540https://github.com/Bing-su/adetailerI believe this is the one everyone uses.
Based helpful anons. We take kindly to newfriends here.
If we only had a FAQ or sometime like that
>>8758563GOOD MORNING SIR
>>8758566It's almost midnight you fucking dumbass
>I'm retarded based on the time of day, anonOh alright.
>>8758568You don't???
wow, so helpful!
>>8757832osakana \(denpa yun'yun\), onyhakase, (jp06, denhijou niki,:0.95) (fellatrix,:0.9) <lora:gopro.safetensors>Weight interpretation style: comfy++Model: bluvolls rectum fantasia (in D minor)>>8757896Pretty much. Him or tree, idk. Not baking the 3B model (no matter how shit he thought it would end up) was an asshole move.
>>8756525you are the most pathetic loser to have ever graced this threadalso the grifter makes lora that can work without being gigantic fat pos
>>8758596>fellatrix??????????
>>8758601Just throwing shit at the wall until something sticks, usually half the shit on that wall doesn't even contribute anything but I can't be bothered to verify.
>>8758601*smacks lips*shiiiiiiiiiiet.
>>8758599it's noodlekek
why is chat kvetching about filesize in 2025?are they poor and in a third world country?
ruror
>>8758637>>8758667>>8758671>>8758673>>8758674samefag
>>8758596box please
>>8758678>>8758680retard>>8757006can you tell what did you change? i'm also getting way too strong colors with my lora using these optionsflow_model = trueflow_use_ot = trueflow_timestep_distribution = "uniform"flow_uniform_static_ratio = 2.5vae_custom_scale = 0.1406vae_custom_shift = -0.4743
tell me name of this sex position?not cowgirl, not reverse cowgirl
oekaki
I came here to begSomeone please do a Tekken style loraPreferably Tekken 5 but 8 is fine too
beg
I beg
on your knees
Let's not be hasty
I fucking hate nai and its shills
>>8758599I ran a dataset through my shitty 1dim config just4ubb. And even compared it directly to his.https://files.catbox.moe/ob3ee8.pnghttps://files.catbox.moe/hq7j9m.pnghttps://files.catbox.moe/fvcikl.pnghttps://files.catbox.moe/qwrk3a.pnghttps://files.catbox.moe/aw3x41.pngbasic overfitment test copy/pasting captions from each respective datasetmine: https://files.catbox.moe/57js0o.pngthe image with that caption: https://files.catbox.moe/difj3q.pnghis: https://files.catbox.moe/l524qv.pngthe image with that caption: https://files.catbox.moe/mjolcr.png (yes this has metadata and yes it uses my loras, just with his godawful shitmix)both are shit. Mine suffers pretty badly of "#yolo, not going to run this twice so if LR and step count is off so be it" syndrome on top of being absolutely starved of dimension. His suffers from being absolutely fried to the point that I can't actually tell if it's legitimately producing NaN's or if the NaN's are just baked in through his garbage dataset since both are entirely viable.if you want my 1dim garbage version: https://www.mediafire.com/folder/y837nnlw31d8jtrained on EPS(had to be for 1:1 apples to oranges), use firedotinc as the activation tag.I don't really recommend using it, though. but you can do whatever, I'm not your dad.At least it's only 5.36MB, though. That's literally 1/20th the size of his!
>>8758678what's wrong with samefagging?
some pre-hallowing shitposting, since nothing going on here
this can't even be considered somewhat anime-related
>>8758751it depend whether it's based on original or remake because konami is anime but poles are /aco/
>>8758178https://files.catbox.moe/q20wpi.png>>8758687Try --cfm_lambda with 0.02
>>8758724oh I forgot to do upscaling tests at all. silly me.in fairness to both sides, upscale is of random old lowres things that used my actual mix(which in theory both are based off of, one even has raw output of it in the dataset!), of which compositions weren't used in either dataset.anyway;https://files.catbox.moe/320ii0.pnglowres used https://files.catbox.moe/vhoajz.pnghttps://files.catbox.moe/su6ixm.pnglowres used https://files.catbox.moe/hin97u.pnghttps://files.catbox.moe/4dwetb.pnglowres used https://files.catbox.moe/6h7z08.pngrandom bonus using a lowres generated from the other lora, for... fairness or somethinghttps://files.catbox.moe/1m20xd.pnglowres used https://files.catbox.moe/4ymrv9.pngand then again but this time generated from minehttps://files.catbox.moe/pbzour.pnglowres used https://files.catbox.moe/udaqv6.pngreally entertaining how a "Cute Simplified Anime Style" lora that is absolutely fried still somehow starts to break down and lets AI realism slop detailing to bleed through when things start getting pornographic. That said, mine actually handled it a lot better than I was expecting. Still suffers pretty badly from being dimension starved, which hurts style reproduction a good bit, but it's at least keeping itself consistent.there isn't really much of a point here. I'm just beating a dead horse.
Even using a regional prompter I am having issues with bleed sometimes....
>>8758781What are you trying to do?
>>8758782Honestly I am just experimenting. One thing I was attempting to do was to get one girl to carry another but that didn't turn out well.
>>8758783Princess carry, over the shoulder, piggyback ride?In my experience regional prompter works best to give characters different features in a prompt that already produces the composition you want without regional. "placing" characters in a specific area works much better imo when combined with controlnet. You can just give it a rough doodle or a child's crayon drawing, and then set up your regions based on where you now know the characters will end up.
>>8758786Princess carry. And I will be sure to keep that in mind, thanks.
>>8758783There is a few positions that are more than doable with pure prompt but you will need to commit to a certain degree of inpaint regardless of what do you intend to use to get the composition like you want>https://files.catbox.moe/q37sor.png>https://files.catbox.moe/97md4u.png
>>8758756>Try --cfm_lambda with 0.02didn't change much unfortunatelyactually i like the bake cfm lambda a tad less
>>8758756thanks buwo.
>>8758756>Try --cfm_lambda with 0.02For what is that? Comfy?
Is it just be or are hands in that bluv*ll goyslop absolutely cooked?
>>8758949https://github.com/bluvoll/sd-scripts
>>8758961Oh, its a sd-scripts fork solely for rectal flow, i see.So anons are actually wasting time with this snake oil? I really cant see any improvements there. If there are, then they must be marginal
>>8758599Beware you will end on pillarinc list
>>8758524>>8758531Adetailer is a band-aid but it works well. Been messing with my artist mixes more and tuning them and honestly with illustrious if you get it right you can forego adetailer entirely for the most part like the world of difference with a few tweaks or including certain artists does wonders for the clarity.
you were laughing at him, calling him a grifter, a failure, an idiot and retard, a jeet, a tasteless slopper. but he didn't give up.look at him now. he continued autistically baking, learning new things, doing experiments, he dropped an unironical sota finetune of noob after almost a year of no one doing anything except lora shitmixing and drowning in snake oil, he essentially dropped noob v2, as a funny test baked on his two 4090s, unironically striking a nerve of our resident shitposter, all while not even visiting this place. he open sourced the code he used to train noob, added support for lora training in it, added new features. all to make training more accessible for people who only have 8gb gpus.what an absolute chad.
tldr?
>>8758992about what?
>>8758688ai ramen
>>8758988whom, what, and where?
>>8759049rectal flow
Yawn, i'm waiting for minthy's 16ch VAE finetune!
>>8758090Do you have any examples? It'd be great to see what people can do with basic models out of the box with the optimizations you mentioned (and what optimizations did you mean?)
What's the most optimal workflow for getting the best quality with the least effort?Gen->Upscale->Adetailer? That seems too basic but it seems to work better than anything else I've tried.
>>8759083Just replace adetailer with manual and proper inpaint and yes, that's all what you need
>>8759084Ah dammit, I suck at inpainting. But thanks though, I do appreciate it.Any advice for getting better at inpainting?
>>8759088Read the furry guide on it. It's more or less good and gives a good grasp of what kind of mentality you need to have while doing so.>https://rentry.org/fluffscaler-inpaint#inpainting-v2
>>8759084missed the least effort part?>>8759083you got the workflow right, now pick a model that's boring and stable thanks to an overpowering style, like WAI>>8759088check the image >>>/e/3014878
>least effortfuck off and kill yourself
>>8759093>now pick a model that's boring and stable thanks to an overpowering style, like WAIWay ahead of you anon!Are there any other models that are good default ones with a decent stable look out of the box?Full size: https://files.catbox.moe/4qazve.png
>>8759098Why are the eyes melting in every single gen?
i don't know how to use bluvoll's snake oil, somebody can share comfy workflow?
>>8759083gen with adetailer > upscalealso use DMD2
>>8759115>DMD2https://github.com/tianweiy/DMD2 ? I've never heard of anyone using this before. It says the study is from 2024. There are a bunch of algorithms like these that are suppose to improve gens but my experience has been they're all inconsistent and not necessarily better than just randomness. Has anyone used this that can vouch for it?
Quick wildcard for the first page of top artists on Danbooru by tag count: https://files.catbox.moe/8brikk.txtFull res: https://files.catbox.moe/y3yteg.png
>>8759119These cousins need to kiss.
https://files.catbox.moe/rhbiqw.png>>8759119for me? it's faster.
>>8759083You should go ask the masters of low effort WAIslop at /edg/, people here will recommend you do stuff like upscale with CN or inpaint, or may even suggest you use models which aren't inherently biased towards butiful detailed 2.8d realism, which are obviously wastes of time.
>>8759083Look up bluvolls rectified flow conversion. You WILL be converted. You will NEVER go back to WAI!! It is perfect.
>>8758850I don't know if there is foolproof check of that in the code but --contrastive_flow_matching is used with it too, have you added that? It is what wrote in the repo
what do you guys use for booru tagging nowadays?Is SmilingWolf--wd-eva02-large-tagger-v3 still the best for it?
Happy Halloween guys!
>>8759381btw, have you ever posted on-topic gen outside of the few gsus gens, like half year ago?at least this time style isn't as ugly as usual
>>8759386Elaborate a bit, was it something you saw on /aco/? Because I have posted some demon girls there too.
>>8759381
>>8759391
>>8759388sorry, mistook you for other anon
please post more 1girl standing
>>8759395And make it WAI anime screencap style.
>>8759395>>8759407samefag
>>8759363>but --contrastive_flow_matching is used with it too, have you added that?damn, no wonder shit didn't workretrained with 0.02 again, but it still doesn't seem to change muchretrained with 0.2, and it's even more saturated than before and there are a lot of artifacts as if i were to raise the cfg
>>8759395like what?
>>8759466after some testing it appears to be a style leakage which somehow doesn't work well with the artist mix i useincreasing the weights of the artist tags fixed the issue
>>8759081Sure, it's cunny though so beware.<https://files.catbox.moe/ruxdgr.mp4>Takes me around 3 minutes to generate at 480p with a 4070 ti. I followed the guide from /ldg/ here: https://rentry.org/wan22ldgguide but picked a quantized model that was under 12 GB. For NSFW it's like SD 1.5 all over again though with a separate lora for everything.
>want more murakami hentai>it just worksthe future is bright
>>8759215>You WILL be convertedI tried and saw no difference with using shitmixes like greek letters.
>>8759080I too can't wait for a VAE that produces greyscale with a shade of blue tint
Any news from the man with the plan and the fursuit and the embezzled money?
>>8759663he quite literally ran away
>>8759658Well it's stable without being a shitmix IS the plus. Though it's not necessarily thanks to rectal flow, more that something was wrong with the LAX's bake that is easily fixed by pretty much any finetune.
>>8759692It's very grey, I don't like it. Feels like eps all over again.
