/hdg/ - Stable Diffusion

Anime girls generated with AI

Index Catalog Archive Bottom Refresh
Name
Options
Subject
Message

Max message length: 12000

files

Max file size: 10.00 MB

Total max file size: 50.00 MB

Max files: 4

Supported file types: GIF, JPG, PNG, WebM, OGG, and more

E-mail
Password

(used to delete files and posts)

Misc

Remember to follow the Rules

The backup domains are located at 8chan.se and 8chan.cc. TOR access can be found here, or you can access the TOR portal from the clearnet at Redchannit 3.0.

US Election Thread

8chan.moe is a hobby project with no affiliation whatsoever to the administration of any other "8chan" site, past or present.

/hdg/ #15 Anonymous 06/06/2023 (Tue) 17:53:35 No. 17518
Previous thread >>16264
>>18702 What's so great about furryshit models that people keep spamming on SD generals? Is it some kind of troll or the models are actually great? I hate furry too much to look into it myself.
>>18703 It's not a troll, they are genuinely better as base models (for furry shit). Better genitalia (infinitely better than NAI derivatives and any lora), more control over the image since way more tags, a lot of fetish shit included, artstyle control by simply using artist names (no need for loras). For /h/ it's not really interesting because even if there is a dick it's not the main focus so people don't care that it's a fleshstick, but for /d/ (and maybe gay shit? idk never went there never will) having a good looking cock is important. I shilled hard for furry models there and eventually people with time and mixing/block-merge skills managed to shit out a few models that drastically improved cock quality. The caveat is that it's a furry model, and it doesn't really like drawing humans so there's still a little wrangling needed to remove the furry part. Seriously, the quality of /d/ before and after furry model merges is night and day. Online service tier disgusting fleshticks in fried images (bajillion loras) vs very good looking cock (with modularity of how you want it to look, foreskin or not, veiny or not, huge or small, thinner base, etc...). Basically good dataset (tags+large sample size) + long but gradual finetuning (model gets better over time but you don't wait a year for the final ver)
>>18704 And how many images are used to fine tune furryshit models?
>>18705 I'm not exactly sure. The fluffyrock project is an immense clusterfuck (see the HF repo https://huggingface.co/lodestones/furryrock-model-safetensors/tree/main), but I believe it was 415k images (but there seems to be a "3M" project started? idfk) as seen on https://civitai.com/models/92527/fluffyrock-pre-3m-autocomplete there's probably more info on the discord but I ain't joining that
>>18706 okay joined it anyway they made a fucking website to help retag shit ( https://tagme.dev/ ) even though it's already quite well why the fuck hasn't anyone done that for anime? Is it just because of the divides between boorus when furry shit is mostly e621?
>>18707 >why the fuck hasn't anyone done that for anime? Is it just because of the divides between boorus when furry shit is mostly e621? Probably because anime already has a usable base model, and the general approach have been to create loras for anything that doesn't work with the base model? Dealing with this fuckton of loras is pretty painful at this point but no one's gonna bother retagging and retraining the whole fucking model. The visual quality is pretty good with anime already so whatever.
>>18708 >anime already has a usable base model NAI was great when it came out, but it was sanitized so much that most shit that isn't 1girl portrait needs a lot of work or is impossible because of the lack of model knowledge. >create loras for anything that doesn't work with the base model Yeah but with that you subject yourself to the inevitable frying, and the fact that loras aren't able to correct huge gaps in model knowledge (for example : cocks). >The visual quality is pretty good with anime already Barely. You can only go so far with models trained on mostly lowres shit. Furries are training on 1088x1088 ffs. And it's not like they've stopped, they're still constantly training when we've been merging loras and block merge inbreeding NAI derivatives for months with little improvement over models like based64.
I just wanted to say that I fucking hate furries and nothing can change that.
>>18710 >furries have their own autistic image tagging operation >meanwhile we're still stuck with NAI and merges and its next iteration will probably be paywalled hard yeah, same
>>18703 >>18704 is keep reading you furry faggots saying this shit and yet nobody posts any proof of it
>>18713 And no one should, furries can go fuck themselves
>>18713 Then go to /trash/ or something you're not finding furfag shit here
>>18714 >>18715 lol i guess holo/seraziel anon isn't welcome here anymore according to you faggots
>>18702 Rome wasn't built in a day and it certainly wasn't built by 10 people either. >>18708 >usable >pretty good To who? Do you seriously not realize just how bad NAI's base model is (particularly the nsfw one) and how many hoops we had to jump through to make it actually usable? Would you like to guess why hands, kemomimi, feet, tails, etc are all fucked and inconsistent?
>>18716 As long as he keeps furryshit to himself (which he does) he certainly is.
>>18709 >most shit that isn't 1girl portrait needs a lot of work or is impossible because of the lack of model knowledge. Yep, anything in landscape mode that isn't an actual landscape is nearly entirely gacha even on the most autistic mixes. This took me a few dozen attempts to gen and I had to look at so much malformed shit. >Yeah but with that you subject yourself to the inevitable frying, and the fact that loras aren't able to correct huge gaps in model knowledge (for example : cocks). Not only that I don't think that guy realizes that if you keep stacking LoRAs you need to: 1) make sure they play nice with each other (and I'm not even talking about noise offset conflicts) 2) make sure they play nice with the base model 3) play the weights game 4) deal with the reduced IT/s There would be so many points of failure.
>>18719 On another note, R-ESRGAN 4x+ Anime6B produces MUCH sharper results than remacri/animesharp/ultrasharp/etc at 0.4, why don't more people use it?
>>18721 I enjoy that when it comes to styles. I don't enjoy that when it comes to concepts.
>>18722 Ah, well that's fair.
>>18720 Damn lmao and here I was thinking if you actually used 6b for that gen. 6b actually does produce sharper lines but it shits on other details and produces visible artifacts. It's almost always obvious if someone used 6b.
>>18724 I don't think 6B is any more prone to getting artifacts than the other upscalers, they're mostly equal in my book. I did notice that 6B tends to eliminate a lot of the blurry and random "details" from hair and such, things you wouldn't classify as artifacts but definitely annoying.
>>18725 Well I guess it's up to your personal preferences but I think 6B is just kinda grating to look at generally. It does have it uses but I personally prefer remacri and sometimes animesharp.
does Hydrus have a way to extract images from Misskey yet? Seems like the place where Cunny artists that got away from twitter are going to upload their stuff now
>>18728 Misskey API uses POST requests which Hydrus does not support currently, you'd need to make some sort of API endpoint yourself and route Misskey URLs thru that. I had to set up something similar for Twitter using gallery-dl and Python's http.server module
>>18730 why does this site bog my browser down so much
New hololiveEN cunny is good and is male collab free, fucking based I want a LORA of her now so I can cum to her jewel feet
>>18732 man i'm glad i don't have your mental illnesses
>>18733 I just dislike male vtubers, dudes are cooler in any other field than girls though
>Set lora to :1 >Doesn't work >Example has it set to -3 >Set it to that >It fucking works How does it work??? Why does this work??? Why do most loras break past 1.5 but this shit works flawlessly at -3???
(2.80 MB 1024x1536 catbox_gh2kib.png)

(2.37 MB 1024x1536 catbox_rj271s.png)

(2.56 MB 1024x1536 catbox_fy66hx.png)

(2.35 MB 1024x1536 catbox_lx8cov.png)

man do i love ai
>>18734 Why would you even watch vtubers, just gen the cuncun
Any new good mixes around or are we still stuck with b64v3?


Forms
Delete
Report
Quick Reply