“…Many of today's problems are a result of yesterday's solutions.” - Otamin
NEURAL NETWORKS ARE THE FUTURE
This tool is a lot better than what we had access to before. It's like DALLE-2 but for practical uses (weeb shit/porn). The specific branch I'm using was trained off of danbooru tags. I recommend you run it while you sleep if you have a slower pc or want to produce large batches to pick through.
GAZE upon randomly generated some 2hu!!!!!!!!!!!
BLEEDING EDGE TECHNOLOGY
also uh some models are now malware watch out hope none of us got pwned
bonus tips for better prompts
we are japanese goblin
sanded out some flaws, it looks pretty good now
>I recommend you run it while you sleep if you have a slower pc or want to produce large batches to pick through.
How? I was running into oom problems if I increased batch size.
Adding flat_chest doesn't flatten chests at all seemingly, you need to put 'breasts' in the negative prompts.
--RUNNING ON 4GB (And under!)--
These parameters are also useful for regular users who want to make larger images or batch sizes!
It is possible to drastically reduce VRAM usage with some modifications:
>Step 1: Edit webui-user.bat
>Step 2: After COMMANDLINE_ARGS= , enter your desired parameters:
>If you have 4GB VRAM and want to make 512x512 (or maybe larger) images,
>If you have 2GB VRAM,
<If you are getting 'Out of memory' errors on either of these, add --always-batch-cond-uncond to the other arguments
If you get a green/black screen instead of generated pictures, you have a card that doesn't support half precision floating point numbers (known problem on 16xx cards):
You must use --precision full --no-half in addition to other flags, and the model will take much more space in VRAM
Make sure to disable hardware acceleration in your browser and close anything which might be occupying VRAM if you are getting out-of-memory errors, and possibly remove GFPGAN (if you have it)
This really hates arms for the most part, other than one non-2hu prompt that came out well. They rarely survive contact with hair.
Can't get doll joints to generate yet
Why does every image with anything that looks kinda like a doll have such an ... expression?
let me help you out fam
((realistic)), [long shot], ((Poorly drawn face)), (Long neck), ((( (cross-eyed )))), ((( blurry ))), ((ugly)), mutation, ((( (mutated hands and fingers )))), deformed, bad anatomy, bad proportions, extra limbs, extra face, ((malformed)), ((neutral expression)), extra_arms extra_legs low_detail bad_anatomy bad_proportions bad_perspective bad_hands bad_feet error
add to negative, shit brix
>taking away bonus limbs
I'll try some of that though, thanks. I was already getting some pretty passable images
you do you, you might be able to get better joints if you try rozen maiden as a franchise and use an artist who draws the dolls exclusively if one exists
I absolutely can't get this shit to generate more than one girl
Was it's training set super fucking limited or something?
I think it's simply due to the fact that solo images are more common and it's hard to squeeze more in the same frame. I got the multiple lolis by adding emphasis and plural tags e.g. ((( multiple_girls ))) 3girls 4girls
AI isn't quite there yet but it would already work for AI assisted images. You can use those images for the general structure and then paint over them and fix all the weak areas.
Seems like a pretty bleak future for anyone that actually likes painting from scratch though.
Browsing is also gonna be hard once you can't discern it from human art easily anymore.
more tips & trix
AI isn't quite there ye-
touhou maribel_hearn with usami_renko blonde_hair brown_hair ((( 2girls ))) multiple_girls puffy_sleeves large_breasts full_body highly detailed trending on artstation thick lines 90s anime artstyle by, wlop shinebell 4k hq high resolution at a black hole in space holding_hands purple_dress dress_shirt purple_eyes
meme, ((realistic)), [long shot], ((Poorly drawn face)), (Long neck), ((( (cross-eyed )))), ((( blurry ))), ((ugly)), mutation, ((( (mutated hands and fingers )))), deformed, bad anatomy, bad proportions, extra limbs, extra face, ((malformed)), ((neutral expression)), extra_arms extra_legs low_detail bad_anatomy bad_proportions bad_perspective bad_hands bad_feet error out of frame
cfg scale 12
ran through img2img and upscaled
You're proofing my point. Hands are broken. Weird clash of artstyles, artifacting around edges, eyes fucked legs fucked. Nothing you can't fix with some touching up though. If you painted over it it would be hard to tell it was generated. img2img is where the real power is at and it's gonna be integrated into most drawing software. There's already tools to do that. See here https://github.com/sddebz/stable-diffusion-krita-plugin
It's mainly the hands. The eyes tend to get more fucked the more parameters you have affecting them. The other issues you brought up can all be resolved with better prompts and running your results through img2img like you said. There's also inpainting for adding specific details but I've had issues with that feature.
WHERE THE FUCK
DO I PUT
WHY DOES NOBODY SAY THIS
export is a bash command that makes a variable visible outside of a script. If you're running a script that ends in *.sh, it should probably go in the start of that file. If you don't know where the script is that's getting run because you're using a UI or something then learn2code fggt
we use windows here
I demand cute hus in sundresses as payment
>bash env variables
pick one nigger
so thats what vagene looks like
I think using a shitton of brackets increases your chances at moeblobs substantially, probably by drowning out the non-booru garbage that's included with the model fnr
tried the newest shit
NovelAI has a thing for images now, there is a leaked model floating around if you can run it and don't want to shill $25, though you get other things for that as well like text gen which can't be run locally. It's also trained on danbooru, as such you will have good luck making the popular characters. As far as I can tell they don't monitor, filter, or otherwise stop you trying to generate no no content; so you're free to generate whatever you want.
Costs anuses to do variations and you can't adjust most parameters on the NAI paid site ver last I checked
All this is pointless till someone makes a trainable model that's not some 40GB vram python abomination since anything you haven't trained yourself with what you want it to actually generate can only generate absolute shit that's just some snapchat filter on some retard 'artist's cowtits whorb
gonno update the OP
so do I just drop the new one in stable diffusion if I bother to find it???
you have a point for making stuff that is less generic looking but some lads are already putting out what I want anyways so I don't mind waiting
Their model is based off of stable diffusion if I remember right, so the leaked model should work. By now there's guides for it somewhere as there are for normal SD and waifu diffusion. Granted I haven't bothered looking in that far, text gen for smut is my bread and butter and their image gen is a nice convenient thing on the side. Cuckchan has some general threads on it, though I haven't been to them in a while since they are a cancer.
wanted a cool anime wallpaper and this absolute faggot uploaded an image for ants with a watermark
so i upscaled it and tried to shop it out with a clean unextended version of the card art *takes breath* pasted over the 4k upscaled watermark and tried to blend the edges together to downsample it again to get something usable
fixed her badge and eyes maybe
these precise edits are aids im basically redrawing shit at this point
The novelai model seems to output higher quality overall when compared to the waifu diffusion model. I've been messing around with editing images using stable diffusion and overall it's been working great. You still have to fix a number of things yourself, however it significantly lowers the difficulty of large edits. The biggest annoyance I've run into is how it has trouble with eyes, hands, and inner layers of clothing. Another thing it struggles with is personality, it's very hard to not get an image that doesn't seem at least somewhat generic.
i KEEP GETTING THE FREAKY EYES AND WEIRD HANDS.