|
Humbug Scoolbus posted:How do you use it/turn it on/prompt it? It might require remix on, but it shows as a third "vary" option next to strong and subtle now when you upscale something
|
# ? Aug 22, 2023 17:31 |
|
|
# ? May 30, 2024 06:12 |
|
|
# ? Aug 22, 2023 20:17 |
|
Happy first birthday Stable Diffusion! Can you believe it's only been a single year since the first Stable Diffusion 1.4 model officially released? No one could have guessed all of the advances we've made. There was a before time, before LoRAs, before ControlNet, before Inpainting, before Txt2Img generation period. Dig through your old folders, what are some of the first things you generated? Let's see just how far things have come in such a short time. What an incredibly wild ride this has been and we're still only just getting started. Make some predictions on what you think the next year will bring! Happy Birthday! Steps: 32, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 1, Size: 512x512, Model hash: 99fd5c4b6f, Model: seekArtMEGA_mega20, Denoising strength: 0.48, Version: v1.2.1, Hires upscale: 1.5, Hires steps: 32, Hires upscaler: 4x-UltraSharp
|
# ? Aug 22, 2023 20:33 |
|
Remaking some of my less successful first stuff for comparisonAARD VARKMAN posted:got dalle 2 access today and after 1 credit I'm already positive I'm going to blow a bunch of money on this
|
# ? Aug 22, 2023 20:40 |
|
KwegiboHB posted:Happy first birthday Stable Diffusion! I think this was DiscoDiffusion "Fear of death" / "Hey Arnold made felt" / "throwing the book at Rudy Giuliani" Fear of Death is the first image I got that made me go "woah" and it's super early in AI development Crayon I think? 1.4s from the discord I know I have more 1.4 stuff someplace. Like I made a peach alien thing where I hosed with a locked seed for several hours figuring out prompt ordering and varying the CFG a bit. e: searched thread found it (8/8/22) pixaal fucked around with this message at 20:57 on Aug 22, 2023 |
# ? Aug 22, 2023 20:48 |
|
With midjourney's new inpainting i was able to go from this: To this: I'm fairly pleased with the results and with how relatively easy it was given the Discord interface on mobile. hydroceramics fucked around with this message at 20:56 on Aug 22, 2023 |
# ? Aug 22, 2023 20:50 |
|
This thread is an impressive documentation of the cambrian explosion of AI image generation over the last two years, particularly the last year
|
# ? Aug 22, 2023 20:59 |
|
Inpainting is so useful (re: MJ) I hosed around with OpenPose XL all day before giving up I did get a good openPoseXL with a turkey I'm probably going to just bin OpenPoseXL because inpainting to fix some stuff up is going to just be way more useful with XL. Not that I couldn't do *both* just I forgot how much I do not understand joints.
|
# ? Aug 22, 2023 21:03 |
|
Thinking back to when the images like this were first going around. https://twitter.com/melip0ne/status/1120503955526750208
|
# ? Aug 22, 2023 21:12 |
|
AARD VARKMAN posted:It might require remix on, but it shows as a third "vary" option next to strong and subtle now when you upscale something You need to have a fully updated version of discord apparently
|
# ? Aug 22, 2023 21:13 |
|
Killer robot posted:Thinking back to when the images like this were first going around. That brown blob is a stuffed teddy bear that got a hole, had the stuffing removed and turned into a scarf. There's a rainbow donkey Pinata behind it. (Feed these into that same AI as part of that prompt and it'd probably clear up a bit)
|
# ? Aug 22, 2023 21:16 |
|
Killer robot posted:Thinking back to when the images like this were first going around. The floor
|
# ? Aug 22, 2023 21:25 |
|
Killer robot posted:Thinking back to when the images like this were first going around. This new era of AI 'adding' to pictures to make them more suggestions of what the original camera took is going to be quite a ride as we settle in. Before when you took a picture you took a picture. Now though, especially with weirdo smartphone apps that is no longer the case, you take a picture and the AI assumes things and adds stuff that wasn't there before. Hope you guys are ready! Prompt was: pile of junk inside a hoarder house piled in the corner on the floor with stained carpet circa 2004 Stable Diffusion XL via crystalClearXL model at .5 denoising, CFG 8, 1024x1025 img2img
|
# ? Aug 22, 2023 22:02 |
|
Here are some of the very first things I created with StableDiffusion. They're honestly not THAT bad, but it's stunning how much the average output quality has improved.
|
# ? Aug 22, 2023 22:29 |
First thing I ever generated. Something from today.
|
|
# ? Aug 22, 2023 22:36 |
Hard to believe this was only May of last year:
|
|
# ? Aug 22, 2023 23:05 |
|
"A pretty lady" See if you can tell which one was midjourney v1 and which one was midjourney 5.2
|
# ? Aug 23, 2023 00:24 |
|
August 15-ish 2022, using Stable Diffusion in Discord before it was publicly released which would have been 1.4, and now with XL: Since this was before the prompt followed the picture in the metadata I had to kinda kludge based on the file name, plus tweak things minorly since how you prompt is pretty different between 1.4/1.5 and XL. No seed anythings either, just img2img at around .5 noise.
|
# ? Aug 23, 2023 00:42 |
|
Same seed image, same prompt...Night in Gotham City 1975
|
# ? Aug 23, 2023 01:52 |
|
The big thing with this one is making the logo using Stable Diffusion, which is pretty neat though it still lacks permanence. The other is her face, character permanence is cool and good. I actually generated a woman's face and am able to carry that through a bunch of different things. Tomorrow I'll try to make the pictures that would actually be in the magazine, though I probably won't bother with the text because a name is one thing, an entire backstory is another. Using AI for productivity rules.
|
# ? Aug 23, 2023 06:22 |
|
Just making my own San Andreas mods in SDXL at this point. render of an angry gandalf riding a bicycle in gta San Andreas, PS2 screenshot, grand theft auto, rockstar games
|
# ? Aug 23, 2023 06:29 |
|
Throwing the book at Rudy on XL (getting strong George Bluth vibes from Latent Rudy) This is much better than the old AI merging his face with a book.
|
# ? Aug 23, 2023 14:48 |
|
|
# ? Aug 23, 2023 14:52 |
|
Killer robot posted:Just making my own San Andreas mods in SDXL at this point. I’m not much of a gamer but the thought of the next GTA (or any game really) leveraging AI character creation is getting me all woozy… truly unique NPCs… create your own player avatar based on a word prompt… imagine the possibilities.
|
# ? Aug 23, 2023 15:17 |
|
Dicty Bojangles posted:I’m not much of a gamer but the thought of the next GTA (or any game really) leveraging AI character creation is getting me all woozy… truly unique NPCs… create your own player avatar based on a word prompt… imagine the possibilities. Making a 3D model vs an image is a bit different, and an optimized 3D model that will actually run well in game without killing performance is completely different. Model optimization is a big field. Now custom textures? Sure but that also already exists in a few games by letting people upload their own for money since you really want that checked. I've gotten some messed up things from Bing (it is very capable of body horror it just does a good job of flagging it most of the time) Now imagine a 12 year old with an entire summer to spend coming up with ways around your filter so they can have a penis shaped hairdo complete with blue highlight in a wavy pattern. pixaal fucked around with this message at 15:29 on Aug 23, 2023 |
# ? Aug 23, 2023 15:27 |
|
Dicty Bojangles posted:I’m not much of a gamer but the thought of the next GTA (or any game really) leveraging AI character creation is getting me all woozy… truly unique NPCs… create your own player avatar based on a word prompt… imagine the possibilities. The 4X strategy game "Galactic Civilizations IV" is using ai in their civ builder for flavor text and custom race artwork, avatars, etc. I know its not quite the same, but as a big fan of the series I'm stoked. The best feature of the series is having a huge array of tools for civ customizing, and the biggest weakness of that system has been the art. I'm very excited to see it at release.
|
# ? Aug 23, 2023 15:51 |
|
pixaal posted:great points Oh yeah totally, my modest mind forgets the potential gore factor. But even if the player can only use a sanitized list of prompts, I imagine that (at least on a subconscous level) the immersive world-building aspect of gaming will undergo a significant shift when game developers start including AI generated skins (and eventually 3d models/fully-formed NPCs). ed.- hydroceramics posted:The 4X strategy game "Galactic Civilizations IV" is using ai in their civ builder for flavor text and custom race artwork, avatars, etc. I know its not quite the same, but as a big fan of the series I'm stoked. The best feature of the series is having a huge array of tools for civ customizing, and the biggest weakness of that system has been the art. I'm very excited to see it at release. Sounds awesome, looking forward to seeing it! Dicty Bojangles fucked around with this message at 15:59 on Aug 23, 2023 |
# ? Aug 23, 2023 15:57 |
|
Dicty Bojangles posted:Oh yeah totally, my modest mind forgets the potential gore factor. But even if the player can only use a sanitized list of prompts, I imagine that (at least on a subconscous level) the immersive world-building aspect of gaming will undergo a significant shift when game developers start including AI generated skins (and eventually 3d models/fully-formed NPCs). Because of the hardware required to run these AI systems (among the most power-hungry things you can currently ask a GPU to even do) it will be a while before the AI systems generate things in real time and more like it's helping generate a big static list of visual bits. So in that way, you won't be making big dongs for your video game whatever in real time, you'll have to pick from things already looked at by human eyes in some form. I can see games doing the real-time thing for NPC conversations but that will probably be done like how current chat systems work now, where you have the customer-facing AI system you are directly interacting with and a second one that does nothing besides police the context between you and the AI system you are interacting with. Even then though, we're going to see static text using AI to generate lines for NPCs before we see the real-time aspect. Prompt policing only goes so far because both the AI systems for visuals and for text are just repeating half-remembered things back to you, the text just the next word and the visual AIs fuzzy concepts you ask it for, by design. It is impossible to tell it to not generate pixels in a specific place in context of whatever else, best you can do is police the training data so when you ask it for a picture of a penis it goes "??? I don't have anything with the tag penis besides this here u go" and what you get doesn't make sense to your human eyes. But the AI did what it does when you ask it for anything else.
|
# ? Aug 23, 2023 16:11 |
|
KakerMix posted:Because of the hardware required to run these AI systems (among the most power-hungry things you can currently ask a GPU to even do) it will be a while before the AI systems generate things in real time and more like it's helping generate a big static list of visual bits. So in that way, you won't be making big dongs for your video game whatever in real time, you'll have to pick from things already looked at by human eyes in some form. I can see games doing the real-time thing for NPC conversations but that will probably be done like how current chat systems work now, where you have the customer-facing AI system you are directly interacting with and a second one that does nothing besides police the context between you and the AI system you are interacting with. Even then though, we're going to see static text using AI to generate lines for NPCs before we see the real-time aspect. Well most moderation is done after generation, so you have another AI look at the image and go "tag everything in this" then when penis shows up with 27% confidence you nuke it from orbit and return Rick Roll.
|
# ? Aug 23, 2023 16:16 |
|
pixaal posted:Well most moderation is done after generation, so you have another AI look at the image and go "tag everything in this" then when penis shows up with 27% confidence you nuke it from orbit and return Rick Roll. Yeah, with the new SD XL though it's ~moderation~ with genitals is because it isn't in the training data to begin with. You can get it to do boobs but that makes sense because boobs in the USA isn't the same as boobs in lots of parts of Europe, so you side-step the issue a bit. With things like MidJourney though, where you are using corporate hardware yeah it does the same thing with the text, a couple of AIs, one you see and the other you don't that is ready to slap down the boner.
|
# ? Aug 23, 2023 16:19 |
|
Dicty Bojangles posted:I’m not much of a gamer but the thought of the next GTA (or any game really) leveraging AI character creation is getting me all woozy… truly unique NPCs… create your own player avatar based on a word prompt… imagine the possibilities. I'm hoping for games where LLM-powered AIs can talk to each other as well as the player, and give rise to interesting emergent behavior. https://arstechnica.com/information-technology/2023/04/surprising-things-happen-when-you-put-25-ai-agents-together-in-an-rpg-town/ That one is a ChatGPT-based research project, but you can run LLMs locally now, even on the CPU (slowly). Pretty sure there's a next-level Dwarf Fortress in there waiting for someone with enough gumption.
|
# ? Aug 23, 2023 16:34 |
|
404 Media, https://www.404media.co/ , which is a bunch of VICE Motherboard writers that bailed when VICE went bankrupt, just released an article titled 'The Community Pushing AI-Generated Porn to ‘the Edge of Knowledge’'quote:On the Discord server for Mage Space, a popular platform for creating AI-generated images, is a list of channels where members share adult content. There are channels for furries, hardcore and softcore porn, and anime. The article is long and proves that this forum is far from the weirdest thing online these days.
|
# ? Aug 23, 2023 16:40 |
|
Humbug Scoolbus posted:404 Media, https://www.404media.co/ , which is a bunch of VICE Motherboard writers that bailed when VICE went bankrupt, just released an article titled 'The Community Pushing AI-Generated Porn to ‘the Edge of Knowledge’' Who thought that we were?
|
# ? Aug 23, 2023 17:43 |
|
well fine be that way XL I found out that XL knows Rudy better than Donald
|
# ? Aug 23, 2023 19:17 |
|
|
# ? Aug 24, 2023 06:15 |
|
Just installed SDXL... it can does text, sometimes, neat!
|
# ? Aug 24, 2023 06:23 |
|
KwegiboHB posted:Happy first birthday Stable Diffusion! 2 years ago I was playing with CLIP+VQGAN... can't find one working anymore, but I liked the abstractness of it. https://i.imgur.com/uDWS2Xy.mp4
|
# ? Aug 24, 2023 06:40 |
|
KwegiboHB posted:Happy first birthday Stable Diffusion! This is back from when I was making stuff on Dall-e mini and being amazed by it.
|
# ? Aug 24, 2023 12:10 |
|
too soon (it's in 2.5 weeks)
|
# ? Aug 24, 2023 13:09 |
|
|
# ? May 30, 2024 06:12 |
|
I'm loving all the early AI stuff people are posting. The evolution is stunning. Although... Even with SDXL and good prompting, sometimes you try for "a woman walking a border collie" and...
|
# ? Aug 24, 2023 15:30 |