|
WhiteHowler posted:This worked for me. thank you thank you I will give this a shot
|
#
?
Sep 10, 2022 19:57
|
|
|
|
| # ? May 29, 2024 01:26 |
|
|
trying out weird/made up artist prompts again a polar bear seated in a restaurant, eating dinner, photographed by graham t. warrenhanger and jonathan banksonian    a polar bear seated in a restaurant, eating dinner, painted by hanson jemmings and gregar the destroyer  
|
|
#
?
Sep 10, 2022 20:56
|
|
|
and actual non-poo poo-post style one i did which i unfortunately can not post the whole prompt for because the file name ends half way through it and i did it days ago e: figured it out, a terrifying statue on a horrible planet, cosmic horror painting, elegant intricate digital painting artstation concept art by Mark Brooks and Brad Kunkle
|
|
#
?
Sep 10, 2022 20:57
|
|
|
drat those are some good bears, AARD VARKMANMercury_Storm posted:Looks like L. Stein implemented the optimized SD scripts into his preloaded dream model now, so if you try to generate larger images than your card has for VRAM it automatically switches to the optimized version. I'm able to generate 1024x1024 images now with 12GB of VRAM on a GTX 3060, though it takes about 3 minutes for one image. Yeah there's a limitation with images larger than 512x512, since that's the size it was trained on. I don't know the details but basically it seems to do larger images in chunks and can lose track of a subject and try to draw it again. But it usually works ok for landscapes and abstract stuff. Doing 512x768 (or vice versa) usually works ok E: Here's an example, the first set was "painting of a beagle" at 512x512, the second set at 1024x1024 with a similar prompt.   mobby_6kl fucked around with this message at 21:11 on Sep 10, 2022 |
|
#
?
Sep 10, 2022 21:07
|
|
|
a giraffe seated at a thanksgiving dinner, eating, painted by one hundred monkeys a giraffe seated at a thanksgiving dinner, eating, painted by jimmy  one thing's for sure: jimmy is not as good a painter as one hundred monkeys but he has a lot of heart
|
|
#
?
Sep 10, 2022 21:26
|
|
|
Ok so if you haven't updated recently, do it now, to the https://github.com/AUTOMATIC1111/stable-diffusion-webui version Negative prompts are maybe the most powerful thing to come out of this, here's the person who made it talking about them: quote:The way this works is different from weights/biases. Stable diffusion sampling works by basically producing a denoised pic with empty prompt, and another pic with same parameters but with desired prompt. Then it looks at difference between them, scales that by CFG scale, and uses that to remove the noise a bit, then the whole thing repeats in a loop. I discovered this by noticing that the empty prompt is wastefully empty and we should put something in there. I initially posted this as Anti-Greg feature, take a picture and make it look like it was not drawn by Greg. There was little interest when I posted about originally a week or 2 ago. putting stuff like: code:*waifu diffusion tag, using negative prompts on that model works REALLY well and can stop some of the more... unique problems in WD: Without negative tags:  With, no other changes: 
Moongrave fucked around with this message at 21:46 on Sep 10, 2022 |
|
#
?
Sep 10, 2022 21:41
|
|
|
I want battlemechs! Not your 'guy in armor' battlemechs, your inverted knee industrial tank style battlemechs. "Long shot" was a pretty useful phrase to keep things in frame, and I had to tell it that they were "huge" and "in a field" or else it would make them look like toys. Here are some of my favorites: This one is close to what I was looking for, but just the cockpit, prompt: photo of huge robot with bird legs, gun arms and helicopter cockpit head  This wasn't really what I was looking for at all, but I love the animalistic battlemech look, prompt: photo a huge battlemech battletech with raptor legs, gun arms and helicopter cockpit head, long shot, standing, reverse joint, claw feet in a field mech  Not great framing, and it leaned way to hard into the bird aspect of things (this happened a lot): photo of huge robot with bird legs, gun arms and helicopter cockpit head long shot standing, birdlike armored legs  I decided to take the last one and do some quick and dirty edits and regeneration. First I edited the top "large robot with cockpit front and cannon on back," then I edited the legs "birdlike armored legs standing on ground"  Okay, that's clearly jumping and not standing... but pretty close. Too bad about the bird that's about to be stomped. The cannon could also be cooler, but at this point I've probably burned through $5 worth of credits so I'm giving it a break.
|
|
#
?
Sep 10, 2022 21:44
|
|
|
BARONS CYBER SKULL posted:Ok so if you haven't updated recently, do it now, to the https://github.com/AUTOMATIC1111/stable-diffusion-webui version Is there a specific way to update outside of pulling the repo down again? I don’t mind doing that, but figured maybe there was a way to run something with an existing deployment. Thank you!
|
|
#
?
Sep 10, 2022 22:02
|
|
|
TwoDeer posted:Is there a specific way to update outside of pulling the repo down again? I don’t mind doing that, but figured maybe there was a way to run something with an existing deployment. Thank you! i just backed up my edited attention.py and dropped the new zip over my old one if you're running hlky's fork from a while back, it's just a flat reinstall yeah
|
|
#
?
Sep 10, 2022 22:06
|
|
|
Nice Van My Man posted:I want battlemechs! Not your 'guy in armor' battlemechs, your inverted knee industrial tank style battlemechs. The fact that it generated chromatic aberration along some of the edges is seriously incredible.
|
|
#
?
Sep 10, 2022 22:10
|
|
|
an extremely fat man from arkansas dressed in american flags. Color photograph by Robert Mapplethorpe and Annie Liebovitz    i like the effect in the first one that crops up every once in a while when forcing color on b&w photographers
|
|
#
?
Sep 10, 2022 22:41
|
|
|
"Thanos painted by George Bush"
|
|
#
?
Sep 10, 2022 22:42
|
|
|
BARONS CYBER SKULL posted:Ok so if you haven't updated recently, do it now, to the https://github.com/AUTOMATIC1111/stable-diffusion-webui version yeah that's a big improvement it's funny that 90% of stable diffusion wrangling is "Make it good." "Don't make it lovely"
|
|
#
?
Sep 10, 2022 23:06
|
|
|
Tunicate posted:yeah that's a big improvement yeah turns out the AI did know how to make hands and non-mutants, we just weren't asking it to not make mutants and hosed up hands
|
|
#
?
Sep 10, 2022 23:22
|
|
|
Bottom Liner posted:The fact that it generated chromatic aberration along some of the edges is seriously incredible. it shows up where it shouldn't on occasion  i've been using the negative prompt to prevent it from occurring and it seems to be working
|
|
#
?
Sep 10, 2022 23:28
|
|
|
drat yeah those negative prompts are excellent. With some extra prompt-smithing on top I asked it or Aerith and it spit this out first try:
|
|
#
?
Sep 10, 2022 23:30
|
|
|
Tunicate posted:yeah that's a big improvement "Great, I'll make sure the chef honors your special request"
|
|
#
?
Sep 10, 2022 23:34
|
|
|
BARONS CYBER SKULL posted:yeah turns out the AI did know how to make hands and non-mutants, we just weren't asking it to not make mutants and hosed up hands It's even closer to a flesh artist then you first thought.
|
|
#
?
Sep 10, 2022 23:40
|
|
|
I'm really happy ai art went open source so soon. I was really worried about it being a corporate/academic only thing.
|
|
#
?
Sep 10, 2022 23:40
|
|
|
messing around with prompt weighting some (meat) balloon!!  lol (this one came from the same batch as the above) 
|
|
#
?
Sep 10, 2022 23:58
|
|
|
     
|
|
#
?
Sep 11, 2022 00:07
|
|
|
owns
|
|
#
?
Sep 11, 2022 00:10
|
|
|
Lol that last one
|
|
#
?
Sep 11, 2022 00:15
|
|
|
BARONS CYBER SKULL posted:faded color manuscript [optional addition to make it less detailed: Codices] (((Anatomy diagram))) of [insert anything you want here], detailed calligraphy writing and symbols, evil runes, bold detailed used this prompt, then put it through SD upscaler and made an animation from the batch https://i.imgur.com/rgfYpvN.mp4
|
|
#
?
Sep 11, 2022 01:08
|
|
|
VectorSigma posted:used this prompt, then put it through SD upscaler and made an animation from the batch oh i do not like seeing this image
|
|
#
?
Sep 11, 2022 01:19
|
|
|
TwoDeer posted:Is there a specific way to update outside of pulling the repo down again? I don’t mind doing that, but figured maybe there was a way to run something with an existing deployment. Thank you! if you download the repo with a git client instead of by just clicking "download zip" on github, you can simply do a git pull and it will automatically grab the changed stuff and fix your local files up to match github has their own graphical git client named "github desktop" which works okay. might still be more trouble than it's worth to learn if you're not otherwise interested in git and programming stuff
|
|
#
?
Sep 11, 2022 02:00
|
|
|
experimenting with the SD upscaler, i've found you can certainly run it without a prompt, but copying the original prompt will give more lifelike results: 
|
|
#
?
Sep 11, 2022 02:02
|
|
|
If you install git, you can put a git link in the web-user.bat in Automatic1111 and it'll auto update for you every time you launch.
|
|
#
?
Sep 11, 2022 02:02
|
|
|
Rinkles posted:What does full ema mean? This is from a few pages ago but "EMA" is in the machine learning model is training data that is only useful if you intend to further train/finetune the model yourself (which takes 30-40gb of vram at minimum, plus of course a dataset). If you just want to generate images, the full-ema model makes absolutely no difference and is just extra filesize for no benefit. e: the "x/y plot" feature is pretty fun to experiment with the differences between different prompts/keywords here's an example with the anime model: 
RPATDO_LAMD fucked around with this message at 03:31 on Sep 11, 2022 |
|
#
?
Sep 11, 2022 02:16
|
|
|
i tried to challenge the negative prompt feature by using the earlier-mentioned bad anatomy string in a composition containing anything touched by rob liefeld. it seems i was correct in my assumption about the difficulty, but i inadvertently created Gandalf the Red while trying to mash up LOTR and Deadpool.
|
|
#
?
Sep 11, 2022 04:08
|
|
|
apparently waifu diffusion cost about $160 usd to fine tune on 50k ish imagea shits getting insanely cheap
|
|
#
?
Sep 11, 2022 04:27
|
|
|
BARONS CYBER SKULL posted:apparently waifu diffusion cost about $160 usd to fine tune on 50k ish imagea One hundred and sixty dollars? lol
|
|
#
?
Sep 11, 2022 04:29
|
|
|
the same price as an 8-month subscription to photoshop (which you just can't buy outright anymore??)
|
|
#
?
Sep 11, 2022 04:46
|
|
|
Never pay Adobe
|
|
#
?
Sep 11, 2022 04:48
|
|
|
KakerMix posted:One hundred and sixty dollars? yeah it was $2 an hour hiring a server farm thing, for 80 hours lmao
|
|
#
?
Sep 11, 2022 09:14
|
|
|
Full training sets are still close to a million though right? Consumer grade 32GB cards will be so much fun in 2-3 years (if they can't lower the VRAM requirements for training)
|
|
#
?
Sep 11, 2022 11:00
|
|
|
pixaal posted:Full training sets are still close to a million though right? Yeah SD 1.4 was $800k Which is still, not a lot
|
|
#
?
Sep 11, 2022 11:02
|
|
|
strange feelings re Daisy posted:"Waiter, I'd like my food to be fresh and delicious please, no garbage scraps thrown in the microwave" The AI is actually one of those rear end in a top hat genies that go "be careful what you wish for, you just might get it mwahaha"
|
|
#
?
Sep 11, 2022 11:30
|
|
|
Kickstarter: train SD on all known internet porn
|
|
#
?
Sep 11, 2022 12:53
|
|
|
|
| # ? May 29, 2024 01:26 |
|
|
Lord Stimperor posted:Kickstarter: train SD on all known internet porn that's the technological singularity
|
|
#
?
Sep 11, 2022 12:56
|
|