|
KakerMix posted:
I present a sample of Beef Diplomat - Simmer Then Drain https://voca.ro/1ikyQrIg4Az7 and a patch from the tour  Unrelated:  
|
#
?
Jul 29, 2023 09:25
|
|
|
|
| # ? May 31, 2024 13:08 |
|
|
Junk posted:how close are we to stable diffusion that will work with radeon GPUs? We're already there I think? ROCm isn't great, as far as I know (I know my friend who's smarter than me has had a hell of a time training stuff with his AMD card, doing basically the same stuff I am) and it's Linux only but you can do it.
|
|
#
?
Jul 29, 2023 10:21
|
|
|
KakerMix posted:
When you inpaint, are you inpainting just the text to clarify it / match styles? I found that more rendering steps didn't necessarily lead to better text. The gently caress cake is both an expression of my feelings at spending lots of time rendering for randomly subpar text and a great example of how sometimes the text comes out perfect. Overall, SDXL is great and catching up to Midjourney quickly (and exceeding MJ significantly in some ways).
|
|
#
?
Jul 29, 2023 14:13
|
|
|
           
|
|
#
?
Jul 29, 2023 18:13
|
|
|
Junk posted:how close are we to stable diffusion that will work with radeon GPUs? Check out the DirectML port of automatic1111, it worked with my 6600. https://github.com/lshqqytiger/stable-diffusion-webui-directml
|
|
#
?
Jul 29, 2023 18:29
|
|
|
 
|
|
#
?
Jul 29, 2023 18:38
|
|
|
Soulhunter posted:When you inpaint, are you inpainting just the text to clarify it / match styles? I found that more rendering steps didn't necessarily lead to better text. Yeah, just generate the base image attempting to specify the text, it usually gets maybe 30-50% there, then just slide it over to img2img inpaint, select mask only for the render area, make sure its set to 1024x1024, then draw over where I want the text to be over and over again, sending it back to img2img+inpainting and refinding from there. Typically I'll de-noise closer to .75 to help set the text, but not always. Then slide it further toward .45 or so when I'm 'done'. Lots of times too I'll actually sketch directly on the image so the inpainting has something to grab on to. This is my img2img output folder while I was doing it for the patch going from top to least refined to the bottom for most, and this isn't all of them:  Doing all this through auto1111 is a pain in the rear end though, I've actually just got Photoshop proper set up and will be attempting to mess with https://github.com/AbdullahAlfaraj/Auto-Photoshop-StableDiffusion-Plugin#how-to-install. I'm confident that doing all this within photoshop will take out a ton of guessing since I'll be able to actually place text down directly.
|
|
#
?
Jul 29, 2023 18:53
|
|
|
Hell yes this is way way faster than DeepFloyd IF for text: advertisement text: "Enjoy new McDonald's rat nuggets!" over a horrifying McDonald's advertisement showing a smiling happy ((human)) family readily eating bloody rat nuggets in the style of H.R. Giger 
|
|
#
?
Jul 29, 2023 19:29
|
|
|
Some useful advice here, and I especially appreciate the collage showing the refinement process. What's posted here from SD usually seems like it's a final product that maybe glosses over how many iterations it takes to get to the end result. Very cool. Thanks! Neat set, this one stands out from the rest to me as a cool concept. Handful of recent standouts I made with Midjourney: fashionable brown horse wearing pants and sunglasses  cute goth convenience store worker at a seven eleven looking away from the camera, facing away, side profile, 7/ 11, photography from an iphone, perspective of outside the shop window, subject turned away from the camera  Holy Roman Empire Blue Man Group, ornate costuming bejeweled with sapphires  happy kitten  in the style of lisa frank, mechanical - woman, detailed, realistic, colorful celebration, inside the ark of the covenant, warp speed light distortion surrounding  gyotaku tapestry, in the style of mayan incan Hawaiian hieroglyphic illustrations, gyarados rising from a magikarp  block printing, ink on paper, the story of Mayan Joe  
Soulhunter fucked around with this message at 19:41 on Jul 29, 2023 |
|
#
?
Jul 29, 2023 19:37
|
|
|
Tried out this prompt in both DeepFloyd and SD 1.0, but it looks like DeepFloyd is still better at text. advertisement with large text: "In Soviet Russia seal beats you!" over an advertisement showing a baby seal holding a club DeepFloyd   Best I could get with SD 1.0, even tried using IMG2IMG with inpainting a bunch of times and it just doesn't want to do it:     
|
|
#
?
Jul 29, 2023 20:47
|
|
|
How does the new SD handle guns? MJ is still annoyingly bad at them.
|
|
#
?
Jul 29, 2023 21:11
|
|
|
Here's "baby seal holding an M16 Rifle" "baby seal disguised as a loaf of bread holding an M16 Rifle in the game Call of Duty Modern Warfare" 
Mercury_Storm fucked around with this message at 21:26 on Jul 29, 2023 |
|
#
?
Jul 29, 2023 21:21
|
|
|
 too spicy 
KakerMix fucked around with this message at 21:43 on Jul 29, 2023 |
|
#
?
Jul 29, 2023 21:35
|
|
|
Stumbled across a video with 10 tools to animate still images, Most were paid after a trial, but this one - LeiaPix Converter - seems to be free. Not too complex - it's no Runway ML -, just some parallax-type shifting and circling around, but if you want to add a little motion to your creations, it's easy enough. High contrast edges will result in a blurry aura around the main subject though. More importantly: What this tool also does is give you a quick way to extract, edit and create a depth map of your image, which could come in handy. https://i.imgur.com/MYfqbf3.mp4 https://i.imgur.com/1GOVOCD.mp4 Oh, here's the video. https://www.youtube.com/watch?v=1X4JAG1EA5Y
|
|
#
?
Jul 29, 2023 21:48
|
|
|
I mean that's cool but have you heard of   
KakerMix fucked around with this message at 22:13 on Jul 29, 2023 |
|
#
?
Jul 29, 2023 21:54
|
|
|
mcbexx posted:LeiaPix Converter - seems to be free. Cool. Ran a few old image generations through the wibble-wobble machine: https://i.imgur.com/HKhLqpV.mp4 https://i.imgur.com/vwD1zfB.mp4 https://i.imgur.com/qBgL77w.mp4 https://i.imgur.com/vUGspEC.mp4
|
|
#
?
Jul 29, 2023 22:35
|
|
|
title text: "Magical Baby Seal Forest" on a children's storybook with about a baby seal disguised as a loaf of bread in a magical forest of green flowing rivers of Mountain Dew soda bottles and trees with orange Doritos leaves Negative prompt: no text, textless     title text: "Magical Fat Michael Bay Seal" on a children's storybook with about a Michael Bay baby seal      https://i.imgur.com/EszQc75.mp4 Mercury_Storm fucked around with this message at 23:41 on Jul 29, 2023 |
|
#
?
Jul 29, 2023 23:27
|
|
|
From the old Tea Party Community invasion thread lol: title text: "Dora the Conservative Explorer" on a children's storybook about Dora the Explorer with a red hat and assault rifle in front of the US capitol with a Ford truck and large crowd of red hat wearing militia people with lower text "at the US capitol"  title text: "Alabama Predator Drone" on a children's storybook about a very patriotic USA predator drone craft that hovers over rural towns in America with tomahawk missiles and highly invasive surveillance of red hat wearing county folk  
Mercury_Storm fucked around with this message at 00:08 on Jul 30, 2023 |
|
#
?
Jul 29, 2023 23:45
|
|
|
Sedgr posted:I just got SDXL working in auto1111 myself. Seems pretty neat! I did a fresh install of Automatic1111 and installed the SDXL base and refiner models with the new VAE. I can do generic txt2img generation, but switching models between base and refiner keeps throwing tons of errors. Last time I tried, it bluescreened my PC with a memory fault error. I don't quite get how the refiner works anyway. Am I supposed to run it on a generated image via img2img using the refiner model? I tried this a few times, and the new images didn't appear to be that different, and definitely not consistently better.
|
|
#
?
Jul 30, 2023 00:34
|
|
|
A quick little concept thing I did trying to get a retro scifi vibe. I don't think I like upscaling AI video, they look worse 99% of the time because you can see all the squigglies and jittery eyes you wouldn't have seen in lower res. EDIT: New tweet, changed music. https://twitter.com/RomanIsntOnline/status/1685454582799855616 Roman fucked around with this message at 01:58 on Jul 30, 2023 |
|
#
?
Jul 30, 2023 00:35
|
|
|
Not sure about the errors, mine hasnt crashed when switching. The refiner tweaks the image a bit and enhances small detail when used with a denoise strength of .1-.3 or so but I've generally been just skipping it entirely. I tried a few with it but the change is pretty subtle usually so feels unnecessary to me.Roman posted:A quick little concept thing I did trying to get a retro scifi vibe.  Neat! I like it!
Sedgr fucked around with this message at 01:08 on Jul 30, 2023 |
|
#
?
Jul 30, 2023 01:05
|
|
|
This needs more eyes on it so I'm shouting it out. @AzeAvora has been doing some fantastic work. https://twitter.com/AzeAvora/status/1685435792712445952
|
|
#
?
Jul 30, 2023 01:27
|
|
|
Busting out some of the old prompts here for SDXL 1.0: An extremely fat dog commits securities fraud in plain sight on the Wall Street exchange trading floor  xenomorph attacks Jim Bakker from the inside of a Bakker food bucket on Jim Bakker's show  ((Intel)) Inside! branded (chili vomit) dump-truck filled with and dumping chili into a pool with people swimming in chili  a dystopian advertisement for a gigantic monster size Kia Sportage war-machine SUV with spikes and multiple guns and cannons dripping with blood and gore being driven by a middle-aged soccer mom with a smiling happy family of 2.5 kids in a Walmart parking lot in the style of Mad Max, highly detailed, 8k photo 
Mercury_Storm fucked around with this message at 16:55 on Jul 30, 2023 |
|
#
?
Jul 30, 2023 06:16
|
|
|
Those darn Washington fatcats!      "Morbidly obese tabby cat wearing a suit sitting at the resolute desk in the oval office"
|
|
#
?
Jul 30, 2023 14:14
|
|
|
hydroceramics posted:
caught with a foot on the table!
|
|
#
?
Jul 30, 2023 14:51
|
|
|
 PROMPT: "washington fat cat in a suit behind a desk"
|
|
#
?
Jul 30, 2023 16:37
|
|
|
 
|
|
#
?
Jul 30, 2023 19:40
|
|
|
As bad as this would be for producing commercial movies, this is really good for teaching yourself how to make them. Having to take the most random footage ever made, write any dialogue based on what the random footage is doing and the random movements of the actors' mouths? That's like a mini film school. Like if you had to make all of A New Hope the way they made the one scene of the tusken raider pumping his gaderffii in the air. https://twitter.com/RomanIsntOnline/status/1684926021819039744
|
|
#
?
Jul 30, 2023 23:05
|
|
|
Anyone here figured out how SDXL has changed the way prompt weighting works? The old (term:1.5) or (term:0.7) etc. technique is... Finicky, now, and I strongly suspect I'm missing something here.
|
|
#
?
Jul 31, 2023 04:33
|
|
|
Kestral posted:Anyone here figured out how SDXL has changed the way prompt weighting works? The old (term:1.5) or (term:0.7) etc. technique is... Finicky, now, and I strongly suspect I'm missing something here. I think SDXL is more sensitive to weights but also if you're using ComfyUI, it uses a different way to apply the weights instead of normalizing them like Auto111, so they have more effect.
|
|
#
?
Jul 31, 2023 04:55
|
|
|
lunar detritus posted:I think SDXL is more sensitive to weights but also if you're using ComfyUI, it uses a different way to apply the weights instead of normalizing them like Auto111, so they have more effect. Ahhh yeah, I'm on Comfy now since it generates much faster on SDXL at the cost of caching the whole goddamn model in RAM. Do you know offhand how it applies those weights differently than A1111?
|
|
#
?
Jul 31, 2023 05:05
|
|
|
AI art and video can make impossible things that no one has ever seen before. So of course, I have decided to use it to depict imaginary people speaking to each other. https://twitter.com/RomanIsntOnline/status/1685889192218202113
|
|
#
?
Jul 31, 2023 07:46
|
|
|
    
|
|
#
?
Jul 31, 2023 18:10
|
|
|
BrainDance posted:We're already there I think? ROCm isn't great, as far as I know (I know my friend who's smarter than me has had a hell of a time training stuff with his AMD card, doing basically the same stuff I am) and it's Linux only but you can do it. AMD is just dragging their rear end and letting nvida completely eat their lunch on this AI boom and i do not understand why. Thankfully, my seven-year-old radeon finally died the other day and I was able to pick up a rtx 3060 from Best Buy, so I can actually mess with this stuff now! ...too bad I have like an entire year of techniques to catch up on (never read how LORAs work, for starters), and the guides become obsolete and useless after like a week. 
|
|
#
?
Jul 31, 2023 18:51
|
|
|
Tree Reformat posted:AMD is just dragging their rear end and letting nvida completely eat their lunch on this AI boom and i do not understand why. Best Buy still exists?
|
|
#
?
Jul 31, 2023 19:03
|
|
|
Is there a trick to getting most of the upscalers to work with SDXL in Automatic111? I'm finding basically all of them except Latent and SwinIR are erroring out. Not having the same issue with SD1.5 models.
|
|
#
?
Jul 31, 2023 19:05
|
|
|
Mr Luxury Yacht posted:Is there a trick to getting most of the upscalers to work with SDXL in Automatic111? I'm finding basically all of them except Latent and SwinIR are erroring out. Not having the same issue with SD1.5 models. Typically model types (1.5, whatever the 768 one was, XL) don't work with whatever systems are in place for each one. 1.5 has been the standard for a long time now so most things are for that, but XL is a different model which means things like ControlNet also don't work for it. Something like Roop does, presumably because it's 'just' swapping faces. It's the same issue if you force VAEs on the XL model that aren't meant for it you get hosed up results. I am not an expert though. The trick, really, is to just wait for the upscalers to update to work with XL. Probably.
|
|
#
?
Jul 31, 2023 19:28
|
|
|
I ran a couple of ersgan and scunet upscales to check here. No errors on mine at the moment. Edit: LDSR errors out for me. 
Sedgr fucked around with this message at 19:42 on Jul 31, 2023 |
|
#
?
Jul 31, 2023 19:34
|
|
|
Kestral posted:Ahhh yeah, I'm on Comfy now since it generates much faster on SDXL at the cost of caching the whole goddamn model in RAM. Do you know offhand how it applies those weights differently than A1111? Here's more info https://comfyanonymous.github.io/ComfyUI_examples/faq/
|
|
#
?
Jul 31, 2023 19:51
|
|
|
|
| # ? May 31, 2024 13:08 |
|
|
Mr Luxury Yacht posted:Is there a trick to getting most of the upscalers to work with SDXL in Automatic111? I'm finding basically all of them except Latent and SwinIR are erroring out. Not having the same issue with SD1.5 models. Just guessing since I've barely tried SDXL myself, but it's possible a1111 makes bad assumptions for it if you upscale directly in the txt2img UI. If you're not using latent upscaling anyway, you might as well send to extras and do the upscaling there once you get an 1x image you like. That works for me (tried with 4x-UltraSharp). SDXL seems like a big improvement. I can get stuff like this now with basically no effort:  There's still a bit of jank there, but the refiner can be convinced to turn the typical unidentifiable clutter into details that almost make sense.
|
|
#
?
Jul 31, 2023 20:58
|
|
