|
mcbexx posted:Are there some tricks to consistently assign colors (or other attributes, really) to single objects in an image without getting much overlap or bleed? Yo, try this. Lemme know how it works for you. https://github.com/hnmr293/sd-webui-cutoff Page snipe edit: I can't believe I almost missed the most important lora of all time  dark soups iii zelda cdi style Steps: 32, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 1391610444, Size: 512x512, Model hash: 99fd5c4b6f, Model: seekArtMEGA_mega20, Denoising strength: 0.69, Hires upscale: 1.5, Hires upscaler: Latent KwegiboHB fucked around with this message at 19:12 on Apr 23, 2023 |
#
?
Apr 23, 2023 18:50
|
|
|
|
| # ? May 30, 2024 23:27 |
|
|
mcbexx posted:I can not for the life of me figure out how not to have a color specifier bleed into other parts of an image. If I want a "blue car", a lot of other elements will turn out blue, even if I assign an individual color to those. Latent Couple
|
|
#
?
Apr 23, 2023 21:04
|
|
|
AARD VARKMAN posted:im game for hero jobs as a theme HERO BOUNCERS EJECT MAGA CHUDS FROM VENUES   HERO CATSTONAUTS PERFORM SPACEWALK REPAIRS   HERO (??) BIOLOGISTS BLEND HUMAN DNA WITH GIRAFFE (I could not for the life of me get a good comp of a giraffe pulling a kid out of a tree)  HERO GOONS SHITPOST AND SAVE THE WORLD (???SOMEHOW?????)  
|
|
#
?
Apr 23, 2023 21:29
|
|
|
mcbexx posted:I can not for the life of me figure out how not to have a color specifier bleed into other parts of an image. If I want a "blue car", a lot of other elements will turn out blue, even if I assign an individual color to those. I use in-painting to change colors of specific things, and if I need a color to be real specific I will take a generated image and put it in Photoshop and color over the area with the specific color, then put it back in stable diffusion to help it along
|
|
#
?
Apr 23, 2023 22:00
|
|
|
I use MJ and don't have any suggestions for how to do it in SD, but manually recoloring something is quick and easy in whatever image editor program. I just toss it in Procreate and add a color layer on top, then set its blending mode to Color/Hue/Hard Light/etc depending on what kind of saturation and depth I want to it. Lately I'm leaning towards generating things in B&W in MJ and then manually colorizing them this way, it's generally better at generating clear/crisp images in B&W and it has weird methods of color use anyway.
|
|
#
?
Apr 23, 2023 22:09
|
|
|
Cabbages and Kings posted:
Did it actually understand "Catstonauts"???
|
|
#
?
Apr 23, 2023 22:44
|
|
|
|
|
#
?
Apr 23, 2023 23:12
|
|
|
Gynovore posted:Did it actually understand "Catstonauts"??? no and mostly it didn't understand that cat astronauts would still need helmets to breathe edit:  i stand corrected, the top row is Deliberate prompted just with "catstonaut". The middle is raw SD 1.5; the bottom row is poo poo a Google Image search for the same term turned up in the first page or so.CATSTONAUTS IN THE TRAINING DATA. The third GIS result appears to be a product that is/was sold on Amazon so there you go, Many Such Items Were Analyzed for latent dreams. (Deliberate)     (SD 1.5)   (google image search)     The Deliberate result with the cat in the little rover is pretty adorable though I wish I had saved the seed. Cabbages and VHS fucked around with this message at 23:42 on Apr 23, 2023 |
|
#
?
Apr 23, 2023 23:32
|
|
|
I ran it through the MJ describe function   
|
|
#
?
Apr 23, 2023 23:54
|
|
|
"Linger not amongst the ashes, lest you carry a fire home."   
|
|
#
?
Apr 24, 2023 00:17
|
|
|
KwegiboHB posted:Yo, try this. Lemme know how it works for you. I had never heard of loras before but they are pretty amazing  
|
|
#
?
Apr 24, 2023 00:52
|
|
|
*scrunching my head as far down as I can* "man I fuckin hate Mondays"
|
|
#
?
Apr 24, 2023 01:29
|
|
|
Ok, I like Grimes. https://twitter.com/Grimezsz/status/1650304205981089793
|
|
#
?
Apr 24, 2023 02:55
|
|
|
Chainclaw posted:I had never heard of loras before but they are pretty amazing That they are, now say goodbye to your harddrive space lol. Most of the 144mb LoRAs from civitai are a bit overcooked because the most popular guide on how to make them when they first came out said to, most would be just fine at 30mb or even 6mb. They're still incredibly smaller than the 2gb+ dreambooth checkpoints they replaced. You can fix that by giving them a weight of about .6 or so, :1 is full strength, :.5 is half, :2 is double, etc. You'll need to experiment but that's half the fun. Also they can be used in the negative prompt to remove the effect from your image, for example IShallRiseAgain's  <lora:bodyHorror_v10:1> ends up making perfect humans.This thread is now 1 years old. Things have uh, changed, every so slightly in the last year. A tad, smidgen if you will. If this feels impossible to keep up with that's because it's impossible to keep up with. I'll try anyways and start working on a more updated Stable Diffusion guide for new people since the current first post is suggesting 1.4. Thank you for that write-up anyways Boba Pearl. I don't use Midjourney, DALL-E or, lol, Craiyon. If someone who does use them wouldn't mind typing a few words towards people starting at square one we could turn the first post into a useful resource. Yes I'm asking you to write a free book report.
|
|
#
?
Apr 24, 2023 03:20
|
|
|
We need an AI art thread, and a second AI "what's going on this week" thread
|
|
#
?
Apr 24, 2023 03:34
|
|
|
Are most people in this thread using Midjourney or Stable Diffusion ? I tried Midourney a couple months ago and wasted all my free credits without getting anything I wanted. I've been using SD through Playground.ai and I'm still not getting very good results, but at least I'm not paying for them. My buddy who is neither technically or artistically inclined is constantly showing me crazy good stuff he's creating in Midjourney. I dunno if I'm just bad at prompting or if I'm using the wrong tool. I'm not opposed to paying for MJ, but I'm not going to bother if my results won't be significantly better than SD.
|
|
#
?
Apr 24, 2023 03:42
|
|
|
Midjourney has improved significantly in the last few months, you should get better results than you did before.
|
|
#
?
Apr 24, 2023 03:59
|
|
|
I read somewhere that you can get extra mid journey credits by... Ranking images? Presumably it takes that ranking data and they use it to tune the model/s? That would be a big step forward and probably why it produces such good imagery now
|
|
#
?
Apr 24, 2023 05:06
|
|
|
You can get extra Fast Hours by voting on Ranked Pair matches: https://docs.midjourney.com/docs/free-hours and yeah it has absolutely helped their models leapfrog past the competition in most regards. Worth noting that it's only extra Fast Hours - you still need a MJ sub and you don't need Fast Hours to generate images. You can see the relative wait times of Fast mode vs Relax mode on their status page: https://status.midjourney.com/ Right now it's a ~2m45s queue for Relax mode and an instant queue for fast mode (still takes ~30 seconds to generate the image), so it remains very usable even when you're out of fast hours. Pro tip if you care about fast hours: it's based on overall usage (and your own usage), so check the status page periodically and switch to relax mode whenever the wait is short (because not many people are currently active), save the fast hours for when the waits are high (at that point it can be ~1min for fast or ~6-7 for relax) ALSO if you sign up for MJ and start using it now, make sure to go into /settings and enable MJ Version 5 - I think it still defaults to Version 4 (could be wrong though) deep dish peat moss fucked around with this message at 05:12 on Apr 24, 2023 |
|
#
?
Apr 24, 2023 05:09
|
|
|
Bucnasti posted:Are most people in this thread using Midjourney or Stable Diffusion ? I'm using SD cause it's free and I already have a gaming computer. Midjourney definitely seems way easier to use though. With SD, you have to choose a good model and possibly loras and keep fiddling until you get something good. Edit: I also recently upgraded and it's crazy how much faster a new video card is.
|
|
#
?
Apr 24, 2023 07:01
|
|
|
Hadlock posted:We need an AI art thread, and a second AI "what's going on this week" thread i tried
|
|
#
?
Apr 24, 2023 07:41
|
|
|
I'm using SD. For pure ease of use Midjourney is simpler if you don't mind the annoyance of it being a discord based input.
|
|
#
?
Apr 24, 2023 08:02
|
|
|
Oh, great, thanks!
|
|
#
?
Apr 24, 2023 08:09
|
|
|
       
|
|
#
?
Apr 24, 2023 10:15
|
|
|
Dang, mind sharing this prompt? Looks like a cool-rear end adult swim show I'd watch.
|
|
#
?
Apr 24, 2023 11:43
|
|
|
feedmyleg posted:Dang, mind sharing this prompt? Looks like a cool-rear end adult swim show I'd watch. a demon symbol that stands on the ground with an orange design, in the style of pop art prints, underground comix, grocery art, intricate underwater worlds, isometric, 1980s, rtx --ar 91:51 --chaos 100 --v 5 It'll be hard to get something like this using the above prompt because of chaos. You could maybe try saving the image and then using it with the /describe function   
Crapple! fucked around with this message at 12:13 on Apr 24, 2023 |
|
#
?
Apr 24, 2023 12:05
|
|
|
https://twitter.com/TheSeaMouse/status/1649750151752216577?s=20
LifeSunDeath fucked around with this message at 13:52 on Apr 24, 2023 |
|
#
?
Apr 24, 2023 13:50
|
|
|
Feeling like Link catching all these fairies in bottles.     
|
|
#
?
Apr 24, 2023 14:59
|
|
|
My grandma would have absolutely loved those
|
|
#
?
Apr 24, 2023 15:07
|
|
|
Bucnasti posted:Are most people in this thread using Midjourney or Stable Diffusion ? Stable Diffusion since it's free and fast on a local machine if you have a gaming GPU. On my 8GB 2070 Super I can crank out a 768x768 image in seconds. Upscaling is what takes a while though and where I run into VRAM limitations (I can 2x upscale 512x512 no problem but for 768 most I can get is 1.75x). You also have more style options with all the different models and Loras and whatnot.
|
|
#
?
Apr 24, 2023 15:15
|
|
|
They also occasionally trade in their bottles for bottles.
|
|
#
?
Apr 24, 2023 15:16
|
|
|
Mr Luxury Yacht posted:Stable Diffusion since it's free and fast on a local machine if you have a gaming GPU. On my 8GB 2070 Super I can crank out a 768x768 image in seconds. Upscaling is memory bound? Must depend on the tool you're using. Granted, I still only use the default ERSGAN upscaler of the InvokeAI Web UI, but I have no problems upscaling my default output size of 576x768 4x on an 8GB 2070S, and sometimes I accidentally upscale the resulting 2304x3072 4x again, which takes a while, but still works. Probably because ERSGAN is tilebased (at least the console output seems to indicate the image gets broken down into smaller tiles).
|
|
#
?
Apr 24, 2023 16:01
|
|
|
AI requirements are all over the place. I've been using a 1660 basic from years ago and I've got a 12GB 3060 coming that I think should make things a bit easier. Going from s/it to it/s will be a nice change. Currently a single 25 step 'Euler a' image takes about 1.5 min which still feels pretty good really when you consider a couple years ago this wasn't a possibility at all.
|
|
#
?
Apr 24, 2023 16:18
|
|
|
mcbexx posted:Upscaling is memory bound? I've been using the R-ERSGAN ones and if I exceed certain upscale bounds while it does start to break it down into more tiles it also throws an Insufficient Memory error once the upscaling starts.
|
|
#
?
Apr 24, 2023 16:27
|
|
|
I haven't tried running SD locally but I've got a gaming rig with a 12gb 4080. How do loras work? I'm mostly looking to do art for my RPG campaigns, so having a model trained to do a specific style would be cool.
|
|
#
?
Apr 24, 2023 16:37
|
|
|
Bucnasti posted:How do loras work? In InvokeAI (starting with v2.34), you download the Lora file into a directory and then select the Lora(s) from a drop down list in the Web UI, which will automatically insertthem into your prompt with a default strength of .75. Then you will need to look on the Lora's download page for the proper keywords that "trigger" the effect the Lora was trained on and other tips about usage and strength. That's it. Automatic1111 should handle them similarly, but I haven't used that UI yet.
|
|
#
?
Apr 24, 2023 16:57
|
|
|
oh drat Invoke finally has LORAs? I thoroughly hosed up my stuff by trying to install web1111 after uninstalling Invoke to play with models. Finally time to undo all that. python path changes here i come
|
|
#
?
Apr 24, 2023 17:00
|
|
|
with Automatic 1111 you click the lora button and it shows you every lora in your lora folder you then click the lora and it slaps it into the positive prompt with a weight of 1.0 and you can adjust that very easily. You can somehow setup cover images for your lora too mine all say no preview image. pixaal fucked around with this message at 17:03 on Apr 24, 2023 |
|
#
?
Apr 24, 2023 17:01
|
|
|
|
| # ? May 30, 2024 23:27 |
|
|
today's theme shan be computers
|
|
#
?
Apr 24, 2023 17:02
|
|