|
BARONS CYBER SKULL posted:Negative prompt: [...], fat, obese, chubby, [...] Now show us the thicc neon pyramids Oh no, a snipe Brutal Garcon fucked around with this message at 04:59 on Sep 21, 2022 |
# ? Sep 21, 2022 04:49 |
|
|
# ? May 30, 2024 14:52 |
|
variants of "a (horned child) on an alter, with a (demon raising a knife) behind it. dark colors, moody painting in the style of Fumito Ueda" keep giving me great images that are nothing like what the original prompt asks for. edit: A Couple more: this one is mignola instead of the ICO guy Jonny Nox fucked around with this message at 09:36 on Sep 21, 2022 |
# ? Sep 21, 2022 09:21 |
|
Stable Diffusion. Look computers don't know how anatomy works ok? dancing skeletons chasing a small scotty dog high contrast, moody painting in the style of mignola Steps: 120, Sampler: Euler a, CFG scale: 12, Seed: 3616764846, Size: 512x512
|
# ? Sep 21, 2022 16:43 |
|
Objective Action posted:The easiest way to just try and add a style or character is to use Textual Inversion https://github.com/rinongal/textual_inversion or https://github.com/nicolai256/Stable-textual-inversion_win if you are on Windows. This isn't actually training the model though, its looking at what the model already knows about and seeing if you can use that vector space to construct a vector basis for the new concept and add it to the lexicon. The upshot is you only need ~5 images to train, the downside is if the vector space doesn't span the basis you need for the concept you can't learn it successfully. I receive the following error using any of the files generated via Textual Inversion in the Web GUI: code:
|
# ? Sep 21, 2022 16:57 |
|
If you are using the Rinongal repo their default readme has you use the older LDM model config yaml files, " configs/latent-diffusion/txt2img-1p4B-finetune.yaml ", that define token lengths of 1280. You want to use the ones in the stable-diffusion folder "configs/stable-diffusion/v1-finetune.yaml" instead. That should resolve your issue. Also note if you want to use textual inversion without needing to use your own hardware another option is HuggingFaces concept library (https://huggingface.co/sd-concepts-library) which has a training notebook (https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/sd_textual_inversion_training.ipynb).
|
# ? Sep 21, 2022 17:30 |
|
Jonny Nox posted:Look computers don't know how anatomy works ok? https://www.youtube.com/watch?v=KUXb7do9C-w
|
# ? Sep 21, 2022 17:52 |
|
Finally found a use for --q 5 on MJ, possibly? Seems to allow fancy shading and linework to survive remaster better than --q 2 so far
|
# ? Sep 21, 2022 18:45 |
|
Midjourney is honestly excellent code:
|
# ? Sep 21, 2022 18:57 |
|
found this dude on google maps and cut him out "a man pointing at an enormous pile of sausage gravy"
|
# ? Sep 21, 2022 19:11 |
|
i appreciate the AI considered he is pointing at gravy and made him fatter
|
# ? Sep 21, 2022 19:12 |
|
Rutibex posted:i appreciate the AI considered he is pointing at gravy and made him fatter That likely has to do with the source image not being the same aspect ratio as the target and it squishing it to get the entire image to get it to work. e: his pixels are under 512 in both directions here's a version you can overlay over any 512x512 you want to feed the AI (or transparent background whatever) pixaal fucked around with this message at 19:20 on Sep 21, 2022 |
# ? Sep 21, 2022 19:17 |
|
Rutibex posted:i appreciate the AI considered he is pointing at gravy and made him fatter it makes him fatter like 2/3rds of the time so far lol i also like how it's giving him bigger and bigger hats ("the coolest thing in the universe")
|
# ? Sep 21, 2022 19:19 |
|
pixaal posted:That likely has to do with the source image not being the same aspect ratio as the target and it squishing it to get the entire image to get it to work. it forces you to crop instead of squishing (note missing sandals, i'm redoing the canvas for that rn), it just loves making this dude fat
|
# ? Sep 21, 2022 19:20 |
|
AARD VARKMAN posted:it forces you to crop instead of squishing (note missing sandals, i'm redoing the canvas for that rn), it just loves making this dude fat Haven't messed with Dalle's version of this well I made a 512x512 that doesn't drop any pixels in an edit if that helps you. You can also add stuff to the background most likely. I know with SD throwing in a camo then saying the subject is where that would be used usually gets you a really nice background.
|
# ? Sep 21, 2022 19:22 |
|
AARD VARKMAN posted:found this dude on google maps and cut him out Ugh, his right arm is giving me those Trypophobia shivers really bad. Have some more Mignola skellies instead. no skin to creep the poor viewer out(and I will leave this prompt until Official Skeleton month)
|
# ? Sep 21, 2022 19:23 |
|
pixaal posted:Haven't messed with Dalle's version of this well I made a 512x512 that doesn't drop any pixels in an edit if that helps you. You can also add stuff to the background most likely. I know with SD throwing in a camo then saying the subject is where that would be used usually gets you a really nice background. I've tried a bunch of combos but DALL-E gives you a 1024x1024 space so it kinda looks like poo poo no matter what. thankfully i have a high res version of the photo i just have to mask first to fit that anyway using the one you posted (thank you!) "a man pointing at a dog eating 4-course thanksgiving dinner"
|
# ? Sep 21, 2022 19:28 |
|
midjourney is not the best when it comes to anatomy and more complex action poses. it tries bless its heart, but doesn't quite make it half the time "a girl with long brown wavy hair and girly glasses, sitting on her bed, wearing a frilly dress, playful, adorable, slim, diffused lighting, in the style of gil elvgren and milo manara"
|
# ? Sep 21, 2022 19:30 |
|
I feel like I should post this... http://www.electricsheepcomix.com/apocamon/
|
# ? Sep 21, 2022 19:39 |
|
does anyone else go down this list and just look for styles they want to use? https://rentry.org/artists_sd-v1-4 for artsy boobs BTW.
|
# ? Sep 21, 2022 20:28 |
|
That list of artist styles is great... Superbowl XX from the stands, 1985 Chicago Bears, copper plate engraving, in the style of Gustave Dore
|
# ? Sep 21, 2022 21:12 |
|
"nuclear mushroom cloud over a medieval castle, painting by Ivan Shishkin, photorealistic, highly detailed, hd, hdr, uhd, unreal engine 5, 8k"
|
# ? Sep 21, 2022 21:39 |
|
|
# ? Sep 22, 2022 00:20 |
|
Kent Brockman Hello
|
# ? Sep 22, 2022 00:31 |
|
The Sausages posted:Hello hel helo
|
# ? Sep 22, 2022 01:18 |
|
the "high-res fix" in the AUTOMATIC1111 repo is great. just be sure to go to the settings tab and set an upscaler, or you'll get blurry results at lower denoising strengths.
|
# ? Sep 22, 2022 03:17 |
|
Wonderful thread, I just got into prompting(not sexual), and my favorite thing to ask real artists right now is "what prompt did you use to make this", they don't find it funny tho Anyway, has this been posted? Text to Pokemon https://replicate.com/lambdal/text-to-pokemon first try: Donnie of course Edit: My latest masterpiece Comfy Fleece Sweater fucked around with this message at 04:53 on Sep 22, 2022 |
# ? Sep 22, 2022 04:33 |
|
i have unlocked unlimited shitposting power
|
# ? Sep 22, 2022 05:00 |
|
Comfy Fleece Sweater posted:
I think this is Cat from Red Dwarf
|
# ? Sep 22, 2022 05:53 |
|
VectorSigma posted:the "high-res fix" in the AUTOMATIC1111 repo is great. just be sure to go to the settings tab and set an upscaler, or you'll get blurry results at lower denoising strengths. These are all extremely dope. Strange cat Tupac is less dope and kind of unsettling.
|
# ? Sep 22, 2022 07:19 |
|
|
# ? Sep 22, 2022 11:53 |
|
Are you telling it to make it look like it came out of a 3D printer or something?
|
# ? Sep 22, 2022 13:24 |
|
And but so indeed, my brethren, I say to you, pray to God most holy every eve, and imbibe your necessary nutrients, and you shall never go astray etc etc Hulk Hogan portrait, intense stare, by Franz Xaver Winterhalter, ez
|
# ? Sep 22, 2022 20:29 |
|
Apologies if it was posted recently, but I've been having trouble finding the local Stable Diffusion webui that supported negative prompts. They don't seem to be supported on my current version, and they seem to be a good way to prevent the mutations that take over people in most of my attempts. A cursory google search brings up the "Automatic1111" fork, which has a webUI - is that a good one?
|
# ? Sep 22, 2022 22:05 |
|
Question Time posted:Apologies if it was posted recently, but I've been having trouble finding the local Stable Diffusion webui that supported negative prompts. They don't seem to be supported on my current version, and they seem to be a good way to prevent the mutations that take over people in most of my attempts. A cursory google search brings up the "Automatic1111" fork, which has a webUI - is that a good one? Automatic1111 is the WebUI that many/most people are using. It includes negative prompts, as well as lots of other cool stuff. There's a feature showcase here: https://github.com/AUTOMATIC1111/stable-diffusion-webui-feature-showcase
|
# ? Sep 22, 2022 22:11 |
|
the newest feature is the High res fix:
|
# ? Sep 22, 2022 22:22 |
|
WhiteHowler posted:Automatic1111 is the WebUI that many/most people are using. Does this work with 8 gigs of memory? I had to use an "optimized" build but that in turn didn't use everything...
|
# ? Sep 22, 2022 23:08 |
|
mobby_6kl posted:Does this work with 8 gigs of memory? I had to use an "optimized" build but that in turn didn't use everything... It works with 4
|
# ? Sep 22, 2022 23:20 |
|
mobby_6kl posted:Does this work with 8 gigs of memory? I had to use an "optimized" build but that in turn didn't use everything... Using default settings it works fine with 8 GB of VRAM to create 512x512 images. On my 8 GB 2070 Super, using DDIM or Euler-A with 30 steps takes around 5 seconds per image.
|
# ? Sep 22, 2022 23:49 |
|
Boba Pearl posted:It works with 4 Okay that's impressive I need to stop dragging rear end updating from waifu with 1.4 official I've been using from day 1. I figured we wouldn't see 4GB until mid next year. This running 1.4 uncensored? I'm reading the documentation and that 4GB seems a bit hacky compared to what I was thinking for next year so maybe it is that long to get the full thing running in 4GB. MedVRAM sounds useful do I need to be able to run conditional and unconditional denoising in the same batch? I'm not sure exactly what that means in relation to what features. I topped out a few times on 1.4 would love to just make that impossible.
|
# ? Sep 22, 2022 23:55 |
|
|
# ? May 30, 2024 14:52 |
|
pixaal posted:Are you telling it to make it look like it came out of a 3D printer or something?
|
# ? Sep 22, 2022 23:58 |