|
A couple more cyberpunk portraits that stuck out from the hundreds I must have churned out over the last couple of days.     
|
#
?
Jan 29, 2023 16:45
|
|
|
|
| # ? Jun 4, 2024 16:29 |
|
|
Yeah some of those new models on Civitai are way better than base Stable Diffusion at fantasy stuff. I've tried creating my D&D group's characters in a few of them and the results are much improved from when I tried the same thing last year in base Stable Diffusion 1.4. Pretty excited to see where things go with this stuff over the next year.
|
|
#
?
Jan 29, 2023 20:17
|
|
|
the most exciting development will be being able to run this poo poo on amd windows without a ton of asterisks
|
|
#
?
Jan 29, 2023 21:04
|
|
|
Still plugging away at this style in MJ. The thought behind it is "Space Fantasy where the barriers between digital and physical reality have eroded" This one doesn't really fit with the others but it's cool:  I've been able to get more detailed with scenes:     But also characters are starting to take shape        In particular I think it's getting to a rad place with the little bits of floating glyphs/matrix code. It's been pretty cool seeing this stuff evolve purely based on the AI, I'm still not using any existing artists' names, genres, real media, etc. Everything is done through emotive text prompts that avoid tropes wherever possible to minimize the "human artist" influence. The only 'real' thing used in the prompts is "ASCII Art" which I use in large part because MJ has no idea how to make actual ASCII art. I want the whole thing to look like it's an oldschool traditional Roguelike game, where the ASCII art of the roguelike game wasn't a technical limitation but was literally representative of what the world looks like. deep dish peat moss fucked around with this message at 21:40 on Jan 29, 2023 |
|
#
?
Jan 29, 2023 21:34
|
|
|
mcbexx posted:A couple more cyberpunk portraits that stuck out from the hundreds I must have churned out over the last couple of days. You aren't using InvokeAi and Proton on a Silicon Mac, are you? I cannot get protogen to work because there are absolutely no instructions at all. I got Invoke, I got Proton, I got the converted files for Silicon Macs, but no idea how they are supposed to all fit together.
|
|
#
?
Jan 29, 2023 22:38
|
|
|
Do you know what chatgpt is currently really bad at? Having negative judgments of people. I played a game where I gave it a character to play (just to avoid some of the low level large language model blockades) and then fed some Reddit AmITheAsshole posts in there. Specifically, I gave it some of the insane stuff where literally nobody would agree with the op. The AI understands how that subforum works and is able to answer in that style, but it is just so incredibly nice and afraid to badmouth anyone. It tries to argue that those people aren't assholes, because they were overcome by emotion, or that their pettiness is just to hard to suppress. I understand that these AIs are supposed to have a generally favourable opinion on humans, but come on! On the other hand I hope I didn't train it to perceive these situations as ordinary. Just in case, I did claim that those posts were fictional in case Assistant is just hesitant to badmouth real people. (For all I know, they are made up.) mcbexx posted:A couple more cyberpunk portraits that stuck out from the hundreds I must have churned out over the last couple of days.
|
|
#
?
Jan 29, 2023 22:57
|
|
|
cant cook creole bream posted:
the gpt model is pre-trained, it is not learning from your inputs in real time or w/e. after MS's Tay companies aren't dumb enough to do that again that "knowledge" only sticks around as long as the individual conversation thread it's in does.
|
|
#
?
Jan 29, 2023 23:59
|
|
|
RPATDO_LAMD posted:the gpt model is pre-trained, it is not learning from your inputs in real time or w/e. after MS's Tay companies aren't dumb enough to do that again Less than that! In my experience it starts forgetting what you've already discussed after around 20 exchanges or so.
|
|
#
?
Jan 30, 2023 00:15
|
|
|
Doctor Zero posted:You aren't using InvokeAi and Proton on a Silicon Mac, are you? I cannot get protogen to work because there are absolutely no instructions at all. I got Invoke, I got Proton, I got the converted files for Silicon Macs, but no idea how they are supposed to all fit together. I'm on Windows, can't help with any Mac related issues., sorry. But with InvokeAI, you just plop the checkpoint file in the models folder and inside of InvokeAI, you open the Model Manager and add Protogen to your list? Do you have the latest version of InvokeAI? Just add "configs/stable-diffusion/v1-inference.yaml" to the config box and the path to your vae-ft-mse-840000-ema-pruned.ckpt in the VAE Location field and you should be good to go? I think the latest version of InvokeAI makes an educated guess even and fills all the info for you. mcbexx fucked around with this message at 01:09 on Jan 30, 2023 |
|
#
?
Jan 30, 2023 00:33
|
|
|
mcbexx posted:I'm on Windows, can't help with any Mac related issues., sorry. Yep. I did all that and when I try to load the model it just thinks for a second and then goes back to the model that it was using before.
|
|
#
?
Jan 30, 2023 00:57
|
|
|
Can anyone explain to me what a NaN is? I've been dealing with an issue with automatic1111 (or maybe with diffusers, it's not clear whether the issue is with the models I trained, with just automatic1111, or with xformers actually. I mean it's definitely an issue with automatic1111 since it's broken for a lot of people and works on earlier commits, but that's probably not the whole story.) I posted a bug report on the automatic1111 GitHub and it turned into a big thing well above my level to fix myself. this was my bug report. Automatic1111 img2ing seems to be broken for some specific models (Including the depth models I trained) but it's a lot of stuff and a lot of people with related errors. But I have no clue what a NaN is and what its significance is. So I'm wondering if, even with the older commit where it works, if this is something that's going to affect the quality of my model, if this is something I should figure out in diffusers/if it's xformers too and retrain my models over.
|
|
#
?
Jan 30, 2023 01:05
|
|
|
BrainDance posted:Can anyone explain to me what a NaN is? not a number
|
|
#
?
Jan 30, 2023 01:07
|
|
|
oh poo poo, well yeah that's more straightforward than I thought. Thanks.
|
|
#
?
Jan 30, 2023 01:13
|
|
|
cant cook creole bream posted:Do you know what chatgpt is currently really bad at? Having negative judgments of people. This sounds similar to Hello History, which has a problem with outright lying to present every historical figure in the best possible light. 
|
|
#
?
Jan 30, 2023 01:58
|
|
|
It's not limited to people, none of the chat AIs are capable of expressing negative opinions in general. It seems to be the only way to keep them from veering immediately into racism and bigotry, but hopefully someone figures out a better solution soon because it really hamstrings their use. e: Probably the funniest and saddest example of this is if you ask ChatGPT or whatever for career advice. It will pretty exclusively tell you that you should start practicing art. If you try to explain to it that "do something creative" is not good career advice it will argue with you and insist that all you need to do is practice art, because it's not willing to express anything other than a glowing positive opinion about how AI is definitely not affecting artists' livelihoods. It will get stuck in a loop of AIsplaining to you that art is the best way to make a living. deep dish peat moss fucked around with this message at 02:07 on Jan 30, 2023 |
|
#
?
Jan 30, 2023 02:01
|
|
|
mcbexx posted:A couple more cyberpunk portraits that stuck out from the hundreds I must have churned out over the last couple of days. I got relatively close to the vibe in MJ, I think.    
|
|
#
?
Jan 30, 2023 02:17
|
|
|
Chamale posted:This sounds similar to Hello History, which has a problem with outright lying to present every historical figure in the best possible light. "Would hitler lie? Yes. #wontfix"
|
|
#
?
Jan 30, 2023 04:57
|
|
|
It's a different vibe but I like how this turned out from just throwing a bunch of cyberpunk buzzwords together plus an opener that's been great for delivering photorealistic outputs for me. I'm not sure what to do about improving the background details though for outputs like this. They're almost always a distracting mess. fake edit: actually I messed around with it a bit before posting and this seems like an incremental improvement:  "1998 photo feature article, lost in the Transmetropolitan city , hacker girl, running down a night street, shadowpunk cyberware interface, bioluminescent tattoo, intricate detail, sharp focus, iso 400, athletic build, plain face, stunning portrait, straight lines, 3000ad"-ive: blurry, blurred, cgi, render, ugly, 3d, digital art, drawing, cartoon, game, actual edit: this one came out nicely enough as well: 
The Sausages fucked around with this message at 14:24 on Jan 30, 2023 |
|
#
?
Jan 30, 2023 05:24
|
|
|
Doctor Zero posted:Yep. I did all that and when I try to load the model it just thinks for a second and then goes back to the model that it was using before. Did you try to reset the WebUI in the Invoke AI settings? If so, and it didn't fix things, I also ran into some weird behavior a while back where the WebUI reset alone would not help, and someone on Github suggested to clear the local file storage in the browser's Java console, which ultimately did the trick. I'm not quite certain if this only affects the storage for InvokeAI or all locally stored browsing data for your local host as well, so use at your own risk (if you have other Apps running locally).
mcbexx fucked around with this message at 09:54 on Jan 30, 2023 |
|
#
?
Jan 30, 2023 09:48
|
|
|
I found that using terms like tattered, broken, torn about clothes can be very good for making post-apocalyptic images in stable diffusion.       
|
|
#
?
Jan 30, 2023 17:31
|
|
|
Lord of the rivets posted:I found that using terms like tattered, broken, torn about clothes can be very good for making post-apocalyptic images in stable diffusion. wow, what ckpt are you using here?
|
|
#
?
Jan 30, 2023 17:52
|
|
|
Analytic Engine posted:wow, what ckpt are you using here? Ever since Elon was on SNL as Wario I've been trying to get less flattering Wario costumes on him. Torn kind of helped but what really made it work consistently was "home made wario costume".  It's not as funny all these months later but damnit I finally got a model and a set of words to work right for the concept I'm happy!
|
|
#
?
Jan 30, 2023 18:01
|
|
|
Analytic Engine posted:wow, what ckpt are you using here? The aEros one
|
|
#
?
Jan 30, 2023 18:09
|
|
|
Tree Reformat posted:the most exciting development will be being able to run this poo poo on amd windows without a ton of asterisks I just want to be able to install it on my computer with an .exe without having to figure out a bunch of computer stuff.
|
|
#
?
Jan 30, 2023 18:16
|
|
|
Fitzy Fitz posted:I just want to be able to install it on my computer with an .exe without having to figure out a bunch of computer stuff. https://nmkd.itch.io/t2i-gui
|
|
#
?
Jan 30, 2023 18:21
|
|
|
Lord of the rivets posted:I found that using terms like tattered, broken, torn about clothes can be very good for making post-apocalyptic images in stable diffusion. Man, the apocalypse looks like a really happy time!
|
|
#
?
Jan 30, 2023 18:59
|
|
|
Thank you. This was easy to install, but unfortunately all of the outputs are green squares, so I guess I will still need to figure out some computer things.
|
|
#
?
Jan 30, 2023 19:20
|
|
|
Doctor Zero posted:Man, the apocalypse looks like a really happy time! Even after the bombs drop, when the photographer says cheese, you smile!
|
|
#
?
Jan 30, 2023 22:31
|
|
|
Doctor Zero posted:Man, the apocalypse looks like a really happy time! Yeah, makes them look like post-apocalyptic cosplayers
|
|
#
?
Jan 30, 2023 22:54
|
|
|
Tried to get the Mona Lisa eating a cheeseburger. Yes:  No: 
|
|
#
?
Jan 31, 2023 00:40
|
|
|
hydroceramics posted:Tried to get the Mona Lisa eating a cheeseburger. NO 
|
|
#
?
Jan 31, 2023 00:42
|
|
|
Mac attack burger
|
|
#
?
Jan 31, 2023 01:01
|
|
|
TIP posted:NO Lmao i was so distracted by the other one
|
|
#
?
Jan 31, 2023 01:04
|
|
|
TIP posted:NO I remember watching some late 90s era history channel show about hauntings and one included this horrible little half dog ghost human graphic that haunted me for years, this reminds me of it severely
|
|
#
?
Jan 31, 2023 04:32
|
|
|
TIP posted:NO That looks like the old rich guy in the wheelchair in the Hannibal movie Red Dragon.
|
|
#
?
Jan 31, 2023 04:46
|
|
|
IShallRiseAgain posted:That looks like the old rich guy in the wheelchair in the Hannibal movie Red Dragon. Do you mean the old rich guy in the wheelchair in the Manhunter movie Silence of the Lambs?
|
|
#
?
Jan 31, 2023 04:56
|
|
|
cant cook creole bream posted:Do you know what chatgpt is currently really bad at? Having negative judgments of people.
|
|
#
?
Jan 31, 2023 05:25
|
|
cowboy
|
Lord of the rivets posted:I found that using terms like tattered, broken, torn about clothes can be very good for making post-apocalyptic images in stable diffusion. Doctor Zero posted:Man, the apocalypse looks like a really happy time!        The power of img2img. Protogen Infinity model (lol at their attempt at licensing lol) , prompt a variation of: (subject and what they are doing if anything) post-apocalypse survivor in a burned/bombed-out dead forest/meadow/small town/city, serious expression, tattered, broken, torn about clothes, dirty skin, unhappy, kodachrome photo 1975 Negative prompts of your choosing, see what works. I used: watermark, cropped, makeup, cartoon, 3d, (disfigured), (bad art), (deformed), (poorly drawn), (extra limbs) Original images thrown into img2img using Auto1111, 50 sample steps, Euler a sampling method, 1024x1024 (as the originals), CFG scale default at 7, Denoising Strength at 0.5. I did not using any face upscaling as someone in the thread already mentioned it smoothes out details, especially when the subject is more far away (the original children (?) in front of the car is noticeable for this. Instead, I just didn't do anything to the faces with the single subjects as they are close enough to the camera that they don't need it. For the two sisters and two kids in front of the car I just inpainted only their faces, on at a time, at full resolution. I also inpainted a bit for the bird lady as I didn't like her ungainly finger under the bird. I also prompted 'holding bird' which explains the ejection of the duck. 'kodachrome' is the first popular color film released and easily gives images that look, and I slapped '1975' on the end to further push the desaturated grainy look.
|
|
#
?
Jan 31, 2023 06:34
|
|
|
KakerMix posted:
Wow those are really good. KakerMix posted:'kodachrome' is the first popular color film released and easily gives images that look, and I slapped '1975' on the end to further push the desaturated grainy look. I'll have to try this out when I get the chance.
|
|
#
?
Jan 31, 2023 07:00
|
|
|
|
| # ? Jun 4, 2024 16:29 |
|
|
Lord of the rivets posted:Wow those are really good. Thanks! Your originals are fantastic and I just wanted to see how it would work altering images in a more subtle way. 0.5 denoising on the img2img setting seems to be always a really good place to start, it keeps a ton of the details, especially things like the woman's hand in her pocket, or the buildings in the background. Interestingly on the guy with the medallions I had to say 'with charms and medallions' for it to preserve them, otherwise it striped them out. Likewise I tried adding 'corpse' for the bird lady but instead it made her holding a skeletal infant which is way too dark. Any model should work for this, and you can see what happens when you go more drawn or animated. I can't help myself from attempting photo-realism so I always go there.
|
|
#
?
Jan 31, 2023 07:05
|
|