|
welcome to The Hive                    
|
#
?
Dec 25, 2022 15:28
|
|
|
|
| # ? May 27, 2024 21:30 |
|
|
I got a new M2 Mac - what's the current best and/or easiest way to get Stable Diffusion running?
|
|
#
?
Dec 25, 2022 18:30
|
|
|
Easiest for Mac is probably InvokeAI https://github.com/invoke-ai/InvokeAI They have an installer here that allegedly works fine: https://invoke-ai.github.io/InvokeAI/installation/010_INSTALL_AUTOMATED/ THere are also instructions for setting up your python environment without the installer if you are comfortable using pip or conda. E: Automatic1111 has more features and will also work, but setting it up is kind of a pain.
|
|
#
?
Dec 25, 2022 18:36
|
|
|
MechaObama will save us   
|
|
#
?
Dec 25, 2022 19:13
|
|
|
Rahu posted:Easiest for Mac is probably InvokeAI https://github.com/invoke-ai/InvokeAI I got InvokeAI set up in about an hour. Then maybe another 30 minutes figuring out how to, and then installing additional models. I tried for a whole loving day to get Automatic1111 to work and finally abandoned it because I couldn’t get it running. The only good thing was I learned a lot about how local SD works trying to troubleshoot it. Go for invoke. There are a lot of features not yet available, but what is there is solid. And now for content. I canceled my MJ account but I’m going to cancel the cancellation because SD isn’t quiiiiiite ready to replace it fully yet. SANTA CLAUSE IS COMING TO TOWN!!!!!  HO  HO  HO  HO  HO  HO  Most pics have prompts for names. Last one I added a little sumpin’ 
|
|
#
?
Dec 25, 2022 20:11
|
|
|
Doctor Zero posted:I got InvokeAI set up in about an hour. Then maybe another 30 minutes figuring out how to, and then installing additional models. what OS and where did you run off the rails with Automatic1111? I generally find it more feature rich, even if Invoke might get there, and say what you will about virtualenv it doesn't annoy me any more than Conda shells. At least in windows I think the process was just: * have git installed * have python 3.9 installed * have checkpoint for 1.5 or 2.1 or w/e * clone repository * copy checkpoint(s) and yaml(s) if needed into /models/stablediffusion/ * follow 2-3 steps to create a virtualenv, activate it * edit webuser-ui.bat to set command line args for XFORMERS etc if you need it, but it will run without all that * run webuser-ui.bat * on 1st run, pytorch and a bunch of other stuff is installed and it will take 30secs - several mins to start depending on your network I think I struggled slightly more to get Invoke to work, it's been a minute.
|
|
#
?
Dec 25, 2022 21:13
|
|
|
The big thing is that the mac install instructions are horribly out of date and won't install all the dependencies you need. You need to manually fiddle with getting the right versions and finding everything that is missing, which I wouldn't recommend for someone who isn't familiar with python package management. You also need to pull a special build of the k-diffusion samplers off of github because the original repo has some incompatibilities with running on apple silicon, and then updating it is a pain because the way that repo is run is a set of changes that gets re-applied over any changes that happen in the orignal repo, so 'git pull' will yell at you and not work and you need to do a 'git reset --hard' and then switch back to the main branch. Overall probably not a great experience for non-developers. e: But i still think it is the best way to run SD so if nothing I said above sounds scary absolutely go for auto1111. If it does probably stick to invokeai.
|
|
#
?
Dec 25, 2022 21:33
|
|
|
Feels like Intel is sleeping on the opportunity to turn their Arc cards into AI cards at the moment. Marketing Intel video cards as not "just for games" but as lower priced alternatives with higher memory for production AI work etc seems like a no brainer but probably difficult from a business pivot. Then again they are starting from scratch on the parallel coprocessor front so why not carve out a niche.
|
|
#
?
Dec 25, 2022 21:43
|
|
|
If you want easy installation, I'll once again recommend https://github.com/cmdr2/stable-diffusion-ui
|
|
#
?
Dec 25, 2022 21:45
|
|
|
https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Installation-on-Apple-Silicon worked fine for me on Ventura 13.1 It takes about 10 seconds to spit out a 512x512 image on a 64gb mac studio, probably slower than a powerful PC GPU but it doesn't feel very long considering the amount of processing going on.
|
|
#
?
Dec 25, 2022 21:46
|
|
|
r u ready to WALK posted:https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Installation-on-Apple-Silicon worked fine for me on Ventura 13.1 a 4090 does a batch of fifty in a minute or two, but 10s a pop is completely usable. 768px images take substantially longer
|
|
#
?
Dec 25, 2022 21:58
|
|
|
It can make some really nice and massive Chistmas parades. or cozy and small ones as well  or even cyberpunk ones 
Vlaphor fucked around with this message at 23:14 on Dec 25, 2022 |
|
#
?
Dec 25, 2022 22:57
|
|
|
Rutibex posted:welcome to The Hive Doctor Zero posted:I got InvokeAI set up in about an hour. Then maybe another 30 minutes figuring out how to, and then installing additional models. What were the model/prompt recipes for these? Also holy poo poo at this thread, the video is amazing: https://www.reddit.com/r/StableDiffusion/comments/zrdk60/great_news_automatic1111_photoshop_stable/ I kinda wish I had a Photoshop setup I could actively draw in to try this out, I mostly do my actual drawing and art stuff on my iPad in Procreate. This is crazy.
|
|
#
?
Dec 25, 2022 23:30
|
|
|
Fuzz posted:What were the model/prompt recipes for these? The prompt for the hive pics is just "hive"  I find that midjourney produces the most interesting output when it's given one word prompts, especially ones that have different meanings based on context
|
|
#
?
Dec 26, 2022 00:00
|
|
|
|
|
#
?
Dec 26, 2022 01:06
|
|
|
 some good science going on here
|
|
#
?
Dec 26, 2022 01:33
|
|
|
Rutibex posted:
"laser chemistry" 
|
|
#
?
Dec 26, 2022 02:51
|
|
|
Fuzz posted:What were the model/prompt recipes for these? Variations on “a dark, sinister Santa Claus shape looming above a tiny town. The Santa shape is huge and horrible and reaching over the town, oil painting style, highly detailed” For automatic1111 I am on a Silicon Mac. So those Windows installers won’t help me. At first I was getting some kind of torch error from python, but I figured out that Homebrew didn’t install correctly, so I fixed that and got rid of that error. But then I kept getting this error - "Exception: Error while deserializing header: HeaderTooLarge" and couldn’t ever get it to work. I found some troubleshooting guides and tried everything, but couldn’t get past it so I gave up. I can follow along the instructions and troubleshooting just fine, but my developer experience is ancient, and I don’t know python. So anything that isn’t specifically included will sail right past me.
|
|
#
?
Dec 26, 2022 04:02
|
|
|
Stable Diffusion is terrible for cluttered backgrounds, anything machinery/equipment related seems to do this poo poo, I haven't worked out how to limit it. Doctor Zero posted:I got InvokeAI set up in about an hour. Then maybe another 30 minutes figuring out how to, and then installing additional models. Seconding this. I've switched and it was a little more work to set up for extra models after the install, but the install went perfectly and for beginners starting with the models they recommend it's preconfigured. It wouldn't be hard to make a script/ui to add models even more easily, but if you can get InvokeAI installed and running you'd have the skills to add models via the CLI. Automatic1111 worked for me eventually but it was a kludge job using a fork that didn't work as described, and had been broken by updates twice. I haven't tried the automated installer that someone made and at this point I'm not going to bother. Invoke has a better workflow for people trying to be creative; it's for people who want to put in a little effort to make beautiful art, or actually use Stable Diffusion professionally. Using it will help you understand how Stable Diffusion actually works. Automatic1111 is easier to use if you can get it running, but I get the impression it's better for people mashing hentai models together using magic words without understanding what CLIP means or what a token is. Invoke is also actually open source, and that's not just a concern for the FOSS fanatics. I'm no expert but from the sounds of it Automatic1111 is apparently a minefield of stolen code and undefined licensing and shouldn't even be looked at by developers. Particularly worrying how they are refusing to license it despite declaring an intention to do so months ago; they know and understand the disadvantaged position this puts users in. Invoke has a permissive MIT license. I'm gonna stop recommending that people use Automatic1111 because the lack of licensing is a real liability - especially if you're using AI artgen as part of a business or nonprofit. Speaking of which I hope that by now any professional creator using these tools is skilled and patient enough to edit out all the flaws/artefacts, and wise enough to shut the gently caress up about how they do it. AI art on its own can't replace artists, same as a textile mill can't replace spinners and weavers without factory staff, but producing low-quality digital art is now unskilled labor. I expect that there's gonna be some real hilarious fuckups over the next forever from this. Like the ESL folks using apps to find out what they'd look like as a carton.
|
|
#
?
Dec 26, 2022 06:42
|
|
|
smbc has been doing some comics about AI art which I've enjoyed, to celebrate here's James Bond Eats A Sandwich run through the machine: or as img2img: 
|
|
#
?
Dec 26, 2022 06:43
|
|
|
"There is no antidote. Both sandwiches were poisoned, Mr. Bond."
|
|
#
?
Dec 26, 2022 06:47
|
|
|
the prompt is "living meat" enjoy                 
|
|
#
?
Dec 26, 2022 15:18
|
|
|
The Sausages posted:
James, God of Sandwiches Doctor Zero fucked around with this message at 19:28 on Dec 26, 2022 |
|
#
?
Dec 26, 2022 19:20
|
|
|
Cabbages and Kings posted:what OS and where did you run off the rails with Automatic1111? I generally find it more feature rich, even if Invoke might get there, and say what you will about virtualenv it doesn't annoy me any more than Conda shells. At least in windows I think the process was just: VERY IMPORTANT BECAUSE PYTHON CAN BE SHITTERS HOW IT IS CONFIGURED. The Automatic script WILL create it's own virtual env by itself as long the shell environment has access to python. It's best to let script do it's own thing then forcing it. Making a virtual env manually and then all of the sudden you see another virtualenv can lead to two gigabyte huge virtualenv (especially with CUDA)
|
|
#
?
Dec 26, 2022 20:23
|
|
|
kids these days are so creative!                                      
|
|
#
?
Dec 26, 2022 21:58
|
|
|
 my photo at the University of Tübingen
|
|
#
?
Dec 26, 2022 22:19
|
|
|
Rutibex posted:the prompt is "living meat" enjoy 
|
|
#
?
Dec 26, 2022 22:37
|
|
|
lol thanks ChatGPT i'll keep that in mind
|
|
#
?
Dec 27, 2022 08:00
|
|
|
design a funny magic card that's a sliver Sure! Here's a design for a humorous Magic: The Gathering card featuring a sliver: Name: Sliver McSliverey Type: Legendary Creature - Sliver Mana Cost: {3} Card Text: Sliver McSliverey has all the abilities of every other sliver on the battlefield. Whenever Sliver McSliverey attacks, you may add a "Sliver Dance" card to your hand. (See below for card text.) P/T: 2/2 Flavor Text: "He's the life of the hive, and he knows all the moves." Sliver McSliverey is a Legendary Sliver creature that has all the abilities of every other sliver on the battlefield. This means that as more slivers enter the battlefield, Sliver McSliverey will become more powerful and versatile. Whenever Sliver McSliverey attacks, you may add a "Sliver Dance" card to your hand, which allows you to perform a special dance that confuses and disorients your opponents. This card is meant to be humorous and lighthearted, adding a bit of levity to the game. Name: Sliver Dance Type: Instant Mana Cost: {1} Card Text: Target player must dance the "Sliver Shuffle" until their next turn. They cannot attack or use activated abilities during this time. Flavor Text: "You'll be shakin' your booty for the rest of the game."
|
|
#
?
Dec 27, 2022 16:05
|
|
|
Lol yeah if you ask for unglued cards the AI will spit out stuff like that
|
|
#
?
Dec 27, 2022 16:35
|
|
|
I wonder if it can come up with half decent Paranoia content.
|
|
#
?
Dec 27, 2022 17:10
|
|
|
KillHour posted:I wonder if it can come up with half decent Paranoia content. write a script about a low security clearance orange team being stuck at the bottom of a pit and then thrown a green rope from above. Write about the green team members interacting with the orange team when the orange team is told they will be executed because touching anything colored orange or above is beyond their clearance and considered treasonous.
|
|
#
?
Dec 27, 2022 18:43
|
|
|
KillHour posted:I wonder if it can come up with half decent Paranoia content.  ChatGPT has an extremely good knowledge of roleplaying games, even extremely obscure role-plying systems. I was able to get it to generate characters from the 90s whitewolf game Mage: The Ascension, full stat sheet and lore appropriate backstory and everything. I was able to get it to generate new spell rotes and mages equivalent of magic items. Its clear whoever the nerds are who are feeding this thing, they are making sure it has a steady diet of roleplaying games Rutibex fucked around with this message at 18:56 on Dec 27, 2022 |
|
#
?
Dec 27, 2022 18:53
|
|
|
Rutibex posted:
Someone needs to transcribe the goon Paranoia podcast going on and train the model on it so we can get an endless supply of parallel universes for it.
|
|
#
?
Dec 27, 2022 19:07
|
|
|
KillHour posted:Someone needs to transcribe the goon Paranoia podcast going on and train the model on it so we can get an endless supply of parallel universes for it. If you want roleplaying content I'll give you a super secret tip. ChatGPT is able to translate anything into anything else not just languages. You can tell it "Rewrite these spells from dungeons and dragons as spells from Mage" or "Convert the Support Gems from Path of Exile computer game into Metamagic feats for D&D". You can translate video game items into roleplaying items. You can ask it for the stats of Excalibur from final fantasy 6 remade into a D&D magic item, then ask it for an appropriate price in gold pieces. It is extremely good at translating concepts back to roleplaying content. You can ask for a paranoia adventure featuring rick and morty and the crew from sesame street. I've already spit out about 200 pages of content. Its written an entire book for me. BTW there are hourly limits on chatGPT output
Rutibex fucked around with this message at 19:20 on Dec 27, 2022 |
|
#
?
Dec 27, 2022 19:15
|
|
|
Been messing around with midjourney again. applying architecture styles to things can produce interesting results. photograph of Baroque Commodore 64  photograph of baroque Nintendo 64  photograph of baroque outhouse  photograph of brutalist Nintendo 64  photograph of Brutalist Commodore 64  photograph of brutalist playground  photograph of gothic commodore 64  Photograph of Xbox Atari 2600  photograph of magazine advertisment for Nintendo Commodore 64  cover of paladin fashion magazine  cover of Barbarian fashion magazine  cover of necromancer fashion magazine  Cinematic Still shot of Johnny Bravo in se7en  Cinematic Still shot of Shaggy and Scooby Doo in se7en  Freddy Krueger and Sadako in the style of a romance novel cover  magazine advertisement for dragon steak 
|
|
#
?
Dec 27, 2022 20:10
|
|
|
The Sausages posted:Automatic1111 is easier to use if you can get it running, but I get the impression it's better for people mashing hentai models together using magic words without understanding what CLIP means or what a token is. sorry but I think this is just incorrect. the automatic1111 webui had more features faster, and they are still not 100% at feature parody, though I think Invoke has come a long way since I started messing with it. I think people who want to go real deep with this probably use both, and then some. At this point in time I'd say automatic1111 is better for doing huge batch gen of initial image sets, experimenting a lot, and probably still my choice for inpainting. Invoke AI great for other post processing and outpainting. Not quite apples and oranges; more a spoon and a fork.
|
|
#
?
Dec 27, 2022 20:30
|
|
|
Cabbages and Kings posted:sorry but I think this is just incorrect. the automatic1111 webui had more features faster, and they are still not 100% at feature parody, though I think Invoke has come a long way since I started messing with it. I think people who want to go real deep with this probably use both, and then some. So where's the spork?
|
|
#
?
Dec 27, 2022 20:35
|
|
|
IShallRiseAgain posted:Freddy Krueger and Sadako in the style of a romance novel cover "Seven days was all it took... To find the man of her dreams!"
|
|
#
?
Dec 27, 2022 21:39
|
|
|
|
| # ? May 27, 2024 21:30 |
|
|
Uploaded a photo of my dog and asked it for a Pulitzer-prize-winning version. This is not quite what I had in mind:
|
|
#
?
Dec 27, 2022 21:40
|
|
