|
That's gonna show up on a power metal album cover
|
# ? Sep 26, 2022 01:20 |
|
|
# ? May 30, 2024 01:08 |
Oh good my nightmare fuel has arrived...
|
|
# ? Sep 26, 2022 03:02 |
|
|
# ? Sep 26, 2022 04:40 |
Oops all demons... Going to stop now before I accidentally summon something.
|
|
# ? Sep 26, 2022 05:08 |
|
is there something i can do to get the automatic111 install to save my styles past the current browser session? the function is awesome but having to paste in my favored styles one at a time is annoying since im using my phone primarily (I'm using runpod.io and the stablediffusion template there (automatic111) has a decent mobile UI finally )
|
# ? Sep 26, 2022 15:40 |
|
AARD VARKMAN posted:is there something i can do to get the automatic111 install to save my styles past the current browser session? the function is awesome but having to paste in my favored styles one at a time is annoying since im using my phone primarily I'm on default settings and it saves them locally in styles.csv for me
|
# ? Sep 26, 2022 15:49 |
|
Tunicate posted:I'm on default settings and it saves them locally in styles.csv for me Oh wait it actually does save them, it just doesn't show the list until you add one. I stopped using it because I thought it wasn't saving them, I'm very stupid also I've given up on turning off my runpod when I'm not using it and just accepting the like $5 a day it's costing me because every time I turn it off it never lets me turn it back on with a graphics card again
|
# ? Sep 26, 2022 15:55 |
|
I saw some tumblr post with the phrase, and decided to try... "Autism burnout Barbie", Cel Shading, Detailed and Intricate, Beautiful Lighting, Line Art nailed it here imo This one's just nice, not sure why it almost ignored the prompt, but it's a nice image And the rest, some good ideas
|
# ? Sep 26, 2022 17:27 |
|
Lol the AI associates autism with a my little pony t-shirt
|
# ? Sep 26, 2022 17:48 |
Baphomet...
|
|
# ? Sep 26, 2022 19:15 |
|
Autistic Ken was awful, no good results even worth posting to make fun of, I imagine because a lot more people draw Barbs than Ken. This one didn't work at all either, in case anyone has any suggestions: Depressed Bugs Bunny from warner brothers shooting up heroin into his arm, disheveled, cartoony proportions, Cel Shading, Line Art It didn't get it very much, and it doesn't seem to know who Tweety bird is either Sedgr posted:Baphomet...
|
# ? Sep 26, 2022 19:31 |
|
As sad as it is, Looney Tunes fell the gently caress off. I remember even adults wearing LT merchandise in the 90's, but when Cartoon Network stopped playing those kind of things if fell off pop culture's radar, so the AI not knowing Bugs or Tweety etc makes sense. Also those Baphomet pics are loving choice. Did you have to tinker a lot?
|
# ? Sep 26, 2022 19:34 |
No tinkering these are unedited. Its interesting how changing the resolution changes the composition. Also, for anyone else using the Automatic1111, you can go into the ui-config JSON and change the Batch count max to whatever. I was running into issues because I couldnt do more than one image in a row changing the batch size without hitting a mem issue. But I could change the batch count and do up to 16. So now with the max changed I can do much larger batches. More Baphomet...aspect change meant lots of faces.
|
|
# ? Sep 26, 2022 19:54 |
|
Maybe my favorite thing I've done in SD so far. 768x512 render and then upscaled for big. Prompt: a beautiful woman wearing a dress made of stars and galaxies and nebulae and constellations, full body view, style of hans zatzka, style of cyril rolando I wish I were better at fixing faces. The image is perfect except that her further-away eye looks weird. Sometimes I can fix minor face issues by increasing or decreasing step count with the same seed, but it didn't work here. I also tried drawing a mask over her face and inpainting from the original background, but the results went from "not any better" to "oh god no what is this hell".
|
# ? Sep 26, 2022 20:41 |
|
I know this is the art thread but I figure someone here knows: what's the current state of play for writing AI? The OpenAI content filter isn't as punishing on GPT-3 as it is on D-2 but I've tripped it a few times and I'd rather not lose access to both figuring out where the line is. Is there a cheaper quality service or an SD equivalent that I can run offline (with hopefully less GPU requirements and possibly a GUI of some kind?) I'm not looking to generate college essays or SEO summaries; my use case is cranking the temperature way up for something like dialogue or short sentences, re-running it a bunch, then panning the reams of incoherent gibberish for the nuggets of humor/creativity. Elotana fucked around with this message at 20:49 on Sep 26, 2022 |
# ? Sep 26, 2022 20:41 |
|
Writing is art too, I’m also curious. And I imagine music AI is already out there? I’m second guessing my choice of learning 3d modeling, since these AIs are almost there too. From reading a bit about the concepts, seems like anything you can teach a person, can be taught to an AI Edit:ahh gently caress I had a prompt idea and I lost it. Write down your ideas, kids
|
# ? Sep 26, 2022 20:50 |
|
WhiteHowler posted:I wish I were better at fixing faces. The image is perfect except that her further-away eye looks weird. Sometimes I can fix minor face issues by increasing or decreasing step count with the same seed, but it didn't work here. I also tried drawing a mask over her face and inpainting from the original background, but the results went from "not any better" to "oh god no what is this hell". Check img2img there's a box that says "fix faces" check that one (in Automatic1111, if you are using something else no idea I know Waifu had it always on). Set de noising to like 0.1 this tells it to basically keep the image the same and only make small changes. 0.2 is usually good but you are looking to just fix up the eyes. It shouldn't stray too far from what you have either so a simplified prompt may work better. Try both importing your positive and negative prompts and then try one with just Woman in a dress with beautiful eyes. By mentioning the eyes it should seek them and work on them not touching the rest of the image. This doesn't always work, Do like 6+ images per batch you only need 1 to be what you want or 1 to be better than what you have and working towards it. You may need to loop it a few times. The rest of it is so good this is probably going to be difficult but some combo should get you there. Worst case you can always save it and use it with future tools. Something in the future should be able to fix that up.
|
# ? Sep 26, 2022 20:52 |
Some beefy beasts with a solid in the middle of the pack.
|
|
# ? Sep 26, 2022 20:57 |
|
Elotana posted:I know this is the art thread but I figure someone here knows: what's the current state of play for writing AI? The OpenAI content filter isn't as punishing on GPT-3 as it is on D-2 but I've tripped it a few times and I'd rather not lose access to both figuring out where the line is. Is there a cheaper quality service or an SD equivalent that I can run offline (with hopefully less GPU requirements and possibly a GUI of some kind?)
|
# ? Sep 26, 2022 20:58 |
|
Comfy Fleece Sweater posted:Writing is art too, I’m also curious. music AI exists but isn't very good https://www.youtube.com/watch?v=bkZrSAZEIeY
|
# ? Sep 26, 2022 21:04 |
|
Tunicate posted:music AI exists but isn't very good some of those bars are legit improvements on the original.
|
# ? Sep 26, 2022 21:07 |
|
mobby_6kl posted:When I read about it, I think the GPT model was like 40 gigs so you probably wouldn't be able to run it on a gamer card.
|
# ? Sep 26, 2022 21:16 |
|
pixaal posted:Check img2img there's a box that says "fix faces" check that one (in Automatic1111, if you are using something else no idea I know Waifu had it always on). Set de noising to like 0.1 this tells it to basically keep the image the same and only make small changes. 0.2 is usually good but you are looking to just fix up the eyes. It shouldn't stray too far from what you have either so a simplified prompt may work better. Thanks! I usually leave Fix Faces on if I'm rendering people. In this case, her left eye is such a different color from the rest of her face that img2img leaves it alone unless I set a really high denoising value, which then modifies too many other things. The area is small enough that I could just touch it up manually in GIMP, but I want to learn how to have the AI do it for me.
|
# ? Sep 26, 2022 21:18 |
|
Elotana posted:There's no loving way a text-to-text model requires 40GB of VRAM when text-to-image models like SD only need 8GB to run offline. Maybe if you're training one from scratch, but I'm just looking for an existing one I can run. Or they just haven't put much effort in reducing the RAM requires because it works and it works on their hardware. They aren't trying to sell an exe they are trying to sell a service, the less portable to lower hardware the better.
|
# ? Sep 26, 2022 21:19 |
|
WhiteHowler posted:Thanks! Sometimes you need to break out GIMP or Photoshop, you can fix it up then run a low denoise to patch it up and keep that AI feel. It might have to sit in your waiting on new tools folder if you want to keep it with no outside editing tools.
|
# ? Sep 26, 2022 21:21 |
|
Elotana posted:There's no loving way a text-to-text model requires 40GB of VRAM when text-to-image models like SD only need 8GB to run offline. Maybe if you're training one from scratch, but I'm just looking for an existing one I can run. The top end text models absolutely require more memory than SD. There are smaller models you can run, but they don't really approach the performance of GPT-3 at things like zero-shot tasks, they do get close at sentence completion. I think the biggest fully released text model is GPT-NeoX-20B. https://blog.eleuther.ai/announcing-20b/ I think that one requires about 40GB of ram, and it has about 1/9 the number of parameters that GPT-3 uses.
|
# ? Sep 26, 2022 21:27 |
|
drat, is StableDiffusion just unusually well-compressed?
|
# ? Sep 26, 2022 21:29 |
|
Apparently language is more complex than image rendering.
|
# ? Sep 26, 2022 21:29 |
|
WhiteHowler posted:Maybe my favorite thing I've done in SD so far. 768x512 render and then upscaled for big. I like to take everyone’s prompts then add “bagpiper”, so here’sy beautiful woman bagpiper from huggingface SD Comfy Fleece Sweater posted:And I imagine music AI is already out there? I’m second guessing my choice of learning 3d modeling, since these AIs are almost there too. I don’t really want full AI songs but I’m extremely interested in AI sounds. “Guitar feedback in a dripping wet cave” “avalanche of broken glass and wind chimes”
|
# ? Sep 26, 2022 21:30 |
|
Elotana posted:drat, is StableDiffusion just unusually well-compressed? Yes. It's the first image model of this quality that can run on consumer cards, it's super amazing what they've managed to pull off.
|
# ? Sep 26, 2022 21:36 |
|
Since this page seems to be the thread's demonic phase, here's what I got from "1980's concert poster for Lucifer the Devil Live at the Rose Bowl": The devil looks like a nice guy. Snowy posted:I like to take everyone’s prompts then add “bagpiper”, so here’sy beautiful woman bagpiper from huggingface SD You monster. I love it.
|
# ? Sep 26, 2022 21:36 |
|
TIP posted:I think the biggest fully released text model is GPT-NeoX-20B. OPT-66B released by Meta might be the largest fully released to date. https://github.com/facebookresearch/metaseq
|
# ? Sep 26, 2022 21:38 |
|
pixaal posted:Apparently language is more complex than image rendering. It's pretty wild GPT works as well as it does, it's honestly more impressive to me than SD. Maybe it's because I haven't looked as much into how it works but SD just makes sense to me while the way GPT can write an entire article about a mouse blender is still blowing my mind.
|
# ? Sep 26, 2022 21:40 |
Baphomet continues to deliver... There are more but really its just endless so at some point its just spamming the thread.
|
|
# ? Sep 26, 2022 22:01 |
|
Tunicate posted:music AI exists but isn't very good are there any AI projects that use midi files as input instead of actual audio?
|
# ? Sep 26, 2022 22:23 |
|
|
# ? Sep 26, 2022 22:57 |
|
a geode split open to display a terrifying cosmic horror inside, elegant intricate digital painting artstation concept art by Mark Brooks and Brad Kunkle detailed Steps: 50, Sampler: Euler a, CFG scale: 7.5, Size: 512x640 Took the same seed and the exact same prompt but with more steps. 60 75 100 150 AARD VARKMAN fucked around with this message at 23:12 on Sep 26, 2022 |
# ? Sep 26, 2022 23:02 |
|
Anyone have access to Midjourney's gallery? Need a large version of the model posted by Mrregista in the top contributions section. Would look great framed.
|
# ? Sep 26, 2022 23:11 |
AARD VARKMAN posted:Neat! Really like the third one.
|
|
# ? Sep 26, 2022 23:21 |
|
|
# ? May 30, 2024 01:08 |
|
"Baphomet wearing a dress made of stars and galaxies and nebulae and constellations, playing a bagpipe, full body view, style of hans zatzka, style of cyril rolando"
|
# ? Sep 26, 2022 23:45 |