|
Question for the thread. I am using Stable diffusion through Google Colab, since I'm on intel Mac's, and it's great! but it's not as good as Stable Diffusion locally. I've seen the results here and in other places. My question is two things. One, if I was building a windows machine for AI art, what is the cheapest good specs I could get to run it locally? And would that also be good for most games (I'm not looking for running at the highest end graphics, just more being able to play stuff like Age of Empires 4 or like through the Halo campaigns, and dolphin emulation, that sort of thing). And, if anyone in this thread is running this on M1/2 MacBook pros, how fast is it? If its working fine on the new Mac's, I honestly am so used to working on Mac's I could just convince myself to upgrade to a new laptop in the next year.
|
#
?
Feb 21, 2023 21:59
|
|
|
|
| # ? May 28, 2024 16:53 |
|
|
My girlfriend is playing a goblin in our Pathfinder game and we were having a hell of a time using stable diffusion to make anything resembling a goblin. I wish people included prompts so I could make a less sexually objectified version that actually has a face. I should ask the TG thread for help making fantasy stuff.
|
|
#
?
Feb 21, 2023 22:02
|
|
|
KPC_Mammon posted:My girlfriend is playing a goblin in our Pathfinder game and we were having a hell of a time using stable diffusion to make anything resembling a goblin. I wish people included prompts so I could make a less sexually objectified version that actually has a face. I should ask the TG thread for help making fantasy stuff. Try... making a male goblin? Then mask out and iterate the face with different prompts, might work? Oh boy, I'm starting to think these perverts might be a problem
|
|
#
?
Feb 21, 2023 22:14
|
|
|
LionArcher posted:Question for the thread. I am using Stable diffusion through Google Colab, since I'm on intel Mac's, and it's great! but it's not as good as Stable Diffusion locally. I've seen the results here and in other places. My question is two things. One, if I was building a windows machine for AI art, what is the cheapest good specs I could get to run it locally? And would that also be good for most games (I'm not looking for running at the highest end graphics, just more being able to play stuff like Age of Empires 4 or like through the Halo campaigns, and dolphin emulation, that sort of thing). And, if anyone in this thread is running this on M1/2 MacBook pros, how fast is it? If its working fine on the new Mac's, I honestly am so used to working on Mac's I could just convince myself to upgrade to a new laptop in the next year. Google Colab can run everything you can run locally*. Are you using an old notebook? The checkpoint should be 2.1-768 the free GPU isn't that fast though and if you are going to pay I'd recommend paperspace (and for the love of god make sure you use a free GPU they'll bill you for all the hours if you're running on a $1/hr card) That said I run locally on a 3070 mobile just fine. I think it's easier to fix a cloud based install than my own laptop OS. If something horribly breaks I don't want to have to gently caress around and reinstall windows. Your notebook just need to pull a newer version if you are happy with your current cloud setup. e: *clarifying I mean with Stable Diffusion version and basic image gen, there's probably some limitations if you are trying to integrate directly with image editing software, img2img and inpainting work pixaal fucked around with this message at 22:25 on Feb 21, 2023 |
|
#
?
Feb 21, 2023 22:16
|
|
|
KPC_Mammon posted:My girlfriend is playing a goblin in our Pathfinder game and we were having a hell of a time using stable diffusion to make anything resembling a goblin. I wish people included prompts so I could make a less sexually objectified version that actually has a face. I should ask the TG thread for help making fantasy stuff. Find an image of a warrior woman and use controlnet to make them into a goblin. Here is an example:   I used the prompt goblin warrior woman, 20yo, RAW Photo, sharp, detailed, 256k film still from a color movie made in 1980, good lighting, good photography, sharp focus, movie still, film grain, greenskin, leather armor, sword Negative Prompt: [[wrinkles]], topless, [[[redskin]]] with the canny model for control. The model was Realistic Vision 13. Denoising strength should 9 to 9.5. I then used the lanzcos upscaler with Codeformers.
|
|
#
?
Feb 21, 2023 22:24
|
|
|
KPC_Mammon posted:My girlfriend is playing a goblin in our Pathfinder game and we were having a hell of a time using stable diffusion to make anything resembling a goblin. I wish people included prompts so I could make a less sexually objectified version that actually has a face. I should ask the TG thread for help making fantasy stuff. try https://huggingface.co/0xJustin/Dungeons-and-Diffusion
|
|
#
?
Feb 21, 2023 22:25
|
|
|
KPC_Mammon posted:My girlfriend is playing a goblin in our Pathfinder game and we were having a hell of a time using stable diffusion to make anything resembling a goblin. I wish people included prompts so I could make a less sexually objectified version that actually has a face. I should ask the TG thread for help making fantasy stuff. for getting a real feel try throwing in the negative new, pristine, perfect that alone usually adds plenty of grit but you can add worn, aged, torn, tattered to the positive prompt if you need further degradation of the subjects. In this case it being a person and not an object, you should get it being less horny, but also try in the negative things like horny, sexy
|
|
#
?
Feb 21, 2023 22:27
|
|
|
KPC_Mammon posted:My girlfriend is playing a goblin in our Pathfinder game and we were having a hell of a time using stable diffusion to make anything resembling a goblin. I wish people included prompts so I could make a less sexually objectified version that actually has a face. I should ask the TG thread for help making fantasy stuff. Its fairly easy, at least on midjourney. "A D&D female goblin rogue"   Edit: with ugly added 
hydroceramics fucked around with this message at 22:39 on Feb 21, 2023 |
|
#
?
Feb 21, 2023 22:36
|
|
|
KPC_Mammon posted:My girlfriend is playing a goblin in our Pathfinder game and we were having a hell of a time using stable diffusion to make anything resembling a goblin. I wish people included prompts so I could make a less sexually objectified version that actually has a face. I should ask the TG thread for help making fantasy stuff. Used that model, this technique along with his prompts https://www.youtube.com/watch?v=xMmCZ1EqMVA and got this with maybe a minute of work:  all within stable diffusion. Now me personally I'd go real hard into intricate details and hand painting and all sorts of stuff, get wild with prompts and things, depending on what sorta character you want. The character sheet is pretty slick imho. EDIT That is to say, there is a lot of room for 'goblin' and right out of the gate, even with MidJourney, is going to make 'pretty' woman since that's how this is all geared toward. Using words like 'homely' can help shift it more toward normality, though even the ~goblin~ I have provided is a a bit cute. KakerMix fucked around with this message at 22:44 on Feb 21, 2023 |
|
#
?
Feb 21, 2023 22:41
|
|
|
pixaal posted:for getting a real feel try throwing in the negative Yeah, negative prompts are the key if you're trying to avoid sexualization - here's a few I generated with different models just making sure to put "boobs, nipples, revealing armor, bikini armor" in the negative prompts. Also put "elf" in the negatives since it kept trying to make them elfy. These are just what it spit out with no touch ups or in painting or anything Positive prompt was "female goblin, green skin, angry face, battle armor, battle axe, full body, action shot" and then some normal parameters I use for good pictures    Not saying these are like, perfect or anything but just showing you can get some non-sexy fightin' goblin women
|
|
#
?
Feb 21, 2023 22:43
|
|
|
All of you are amazing, thank you so much for pointing me in the right direction. The SD models I was using can make dwarves and elves just fine but had no idea how to handle goblins. I'll give it another try this evening.
|
|
#
?
Feb 21, 2023 23:22
|
|
|
LionArcher posted:And, if anyone in this thread is running this on M1/2 MacBook pros, how fast is it?
|
|
#
?
Feb 21, 2023 23:59
|
|
|
Ruffian Price posted:on the shittier M1 Max (24c GPU) it takes me around 40 seconds for a 512x512 image, 2-3 minutes for 1024x1024 and 6-7 for 1280x1280 (25 steps DPM 2M Karras) Do you mind checking to see how long a single 2560x1600 image takes?
|
|
#
?
Feb 22, 2023 00:22
|
|
|
KakerMix posted:I know we like to joke around, have a good time, a few laughs in this thread but do not be surprised at what most AI image generation is being used for right now. 
|
|
#
?
Feb 22, 2023 00:32
|
|
|
KPC_Mammon posted:My girlfriend is playing a goblin in our Pathfinder game and we were having a hell of a time using stable diffusion to make anything resembling a goblin. I wish people included prompts so I could make a less sexually objectified version that actually has a face. I should ask the TG thread for help making fantasy stuff. Midjourney absolutely blows SD out of the water for pretty much everything these days, but especially for stuff like this. No special prompting used for these, just "Female Goblin Fighter, D&D portrait" - you can play with it until the hands look right pretty easily.   
deep dish peat moss fucked around with this message at 00:37 on Feb 22, 2023 |
|
#
?
Feb 22, 2023 00:34
|
|
|
It's very funny to me how much these struggle with hands. It's getting better but still.
|
|
#
?
Feb 22, 2023 00:43
|
|
|
Confusedslight posted:It's very funny to me how much these struggle with hands. It's getting better but still. just throw inpainting over the hands write hands, tell it to make 200 copies and mix a drink. You should end up with several usable ones to choose from. It'd be very wasteful to do this with MJ, with SD, you usually are paying by the watt or hour. Just let it rip.
|
|
#
?
Feb 22, 2023 00:45
|
|
|
pixaal posted:just throw inpainting over the hands write hands, tell it to make 200 copies and mix a drink. With control net, you can just control net openpose(hands) a new set of hands, and then do a hed on the whole thing to make it mesh.
|
|
#
?
Feb 22, 2023 00:49
|
|
|
deep dish peat moss posted:Midjourney absolutely blows SD out of the water for pretty much everything these days, but especially for stuff like this. MJ is still amazing, but "playing around with it" with MJ means just generating variations on the same prompt until it finally generates something un-hosed while SD had pretty good inpainting for months now, even before this week's ControlNet breakthrough. I've been doing more and more of "generate something cool looking but hosed in MJ, then fix it in SD" in the last month or two. Here's a good post from a fellow goon demonstrating the workflow: https://forums.somethingawful.com/showthread.php?threadid=4020822&pagenumber=3#post529684622
|
|
#
?
Feb 22, 2023 00:50
|
|
|
Not really, you can literally add "with accurate hands" or "with realistic hands" to the prompt, for one.   Or you can image prompt an existing generation Or change the pose, or the style cues, etc. MJ isn't bad at hands, it's just the "default MJ styling" that can't do hands well... unless you tell it to. The newest model of MJ (released on Feb 7th) is actually really drat good at hands overall other than throwing on a 6th finger a lot of the time. deep dish peat moss fucked around with this message at 01:01 on Feb 22, 2023 |
|
#
?
Feb 22, 2023 00:58
|
|
|
I wish I had 6 fingers, I could give the middle finger to people twice per hand.
|
|
#
?
Feb 22, 2023 01:07
|
|
|
So, I did an experiment and I figured out you can use ControlNet to colorize black and white images really well.  I generated the version with colors, put in a photoshop as a color layer for the original image, and did a very light touch-up on the hand. Here are some bonus images I used the newest Junji Ito anime to generate this  Also some more Vampire the Masquerade NPCs  
|
|
#
?
Feb 22, 2023 01:35
|
|
|
deep dish peat moss posted:Midjourney absolutely blows SD out of the water for pretty much everything these days, but especially for stuff like this. There's a very important thing that Midjourney doesn't even compete with Stable Diffusion on KakerMix posted:Porn. It's porn. also not being open source and free and having a shitload of breaking tech coming out like the ControlNet stuff. But it probably better if you just want a simple "plug and play" solution
|
|
#
?
Feb 22, 2023 01:41
|
|
|
Stable Diffusion has a cheap API available, which is a point in the column for that.
|
|
#
?
Feb 22, 2023 01:50
|
|
|
I let my computer hallucinate on Deliberate v1.1 while I was getting dinner. I had it make 150 images with no prompt at 1280x720, running DPM++ 2S a Karras for 35 steps. The CFG scale was 1 and the seed was random. These are my favorites      The best of the bunch, TIMG due to nudity  Finally, a gallery of them all 
|
|
#
?
Feb 22, 2023 02:54
|
|
|
deep dish peat moss posted:Midjourney absolutely blows SD out of the water for pretty much everything these days, but especially for stuff like this.  MidJourney is great with how little you have to do to get a good image out of it, but it is very rigid, restrictive, and distinctive. MidJourney looks like MidJourney, and if you do want to reach in and fix things you're SOL. At some point AI image generation will be good enough to where people like me won't be able to see the hands or eyes or weirdo extra elbow nub, but it isn't here yet. I'm tool agnostic, and signed up for a MidJourney corporate account right when it became available, so I don't want to give the impression that I'm somehow above it. It's just that I can not help but want to fix the issues I see especially since I consider them so obvious. Bonus ~real~ goblin lady, though she now is sporting some sort of hosed up knuckle blade thing that I don't want to reach in and fix like I did last time. Does show the potential of ControlNet though. 
KakerMix fucked around with this message at 03:21 on Feb 22, 2023 |
|
#
?
Feb 22, 2023 03:08
|
|
|
Trying my hand at the Goblin prompt, still in DeliberateV11.  A little green goblin William Defoe in this one.   All generated using "goblin woman in black leather armor, green skin, red hair, d&d, dungeons and dragons, photo, 4k, raw photo, profile, black background" and negative "nipples, topless, nudity" e: 
LASER BEAM DREAM fucked around with this message at 04:04 on Feb 22, 2023 |
|
#
?
Feb 22, 2023 03:58
|
|
|
LionArcher posted:Do you mind checking to see how long a single 2560x1600 image takes? Doing one that big will probably make you run out of ram unless you spend the big bucks for 64gb+. On my m1 pro mbp I use about 30 of my 32gb to make a 896x1024 image. Going to 1024x1024 makes it start swapping a bunch and kills performance. e: I should add that is for making an image at a 'native' 896x1024, specifically when using the 'latent' upscaling mode in Automatic1111 from 448x512 to 896x2014. Other upscalers can work with less memory but the results won't be as good. And to add more specifics: My specific processor is the M1 Pro with 14 GPU cores and 32GB ram Using the DPM++ 2M Karras sampler which generally needs ~15 samples for a preview and ~25 for a 'final' render At 896x1024 it runs at ~6.1 seconds per iteration. At 512x512 it is 1.2 iterations per second. Rahu fucked around with this message at 04:23 on Feb 22, 2023 |
|
#
?
Feb 22, 2023 04:09
|
|
|
The others have got the realistic look covered but playing around with the various styles while you hold the seeds fixed can give you really different feels. These are all raw cause I didn't want to spend the time upscaling and inpainting them (I just cropped them, lol) but just so you can see what I mean: Fighters:     Gladiators:     Rogues:     Barbarians:      Wizards:         Monks:  
|
|
#
?
Feb 22, 2023 04:28
|
|
|
Doctor Zero posted:Doctor Zero is taken by a Strange Mood. I was inspired to try out some Jim Henson muppet aliens with controlnet      
|
|
#
?
Feb 22, 2023 05:09
|
|
|
|
|
#
?
Feb 22, 2023 05:37
|
|
|
And just because, why not? Animu Gerblins:              
|
|
#
?
Feb 22, 2023 05:38
|
|
|
General Bullshit › AI Art: It is criminal to not post your goblins
|
|
#
?
Feb 22, 2023 05:47
|
|
|
|
|
#
?
Feb 22, 2023 16:29
|
|
|
https://imgur.com/DyOG4Yz While Yudkowski might have some good points, at least we got this really horny poo poo out of this before the inevitable apocalypse.
|
|
#
?
Feb 22, 2023 16:39
|
|
|
KakerMix posted:lmfao Can you change the bike to a Hayabusa
|
|
#
?
Feb 22, 2023 17:03
|
|
|
TIP posted:I was inspired to try out some Jim Henson muppet aliens with controlnet  Tagichatn posted:General Bullshit › AI Art: It is criminal to not post your goblins ok. Midjourney generates the same drat goblins. ("Female Goblin Thief")  So I did something different: ("Goblin Yoga instructor in an upscale yoga studio")  "Photorealistic Female Goblin Footpad"   
Doctor Zero fucked around with this message at 17:25 on Feb 22, 2023 |
|
#
?
Feb 22, 2023 17:22
|
|
|
LASER BEAM DREAM posted:reddit is really gross. This is from the SD sub right now. Flip dat poo poo. Let's do an AI Photoshop Phriday level idiocy event: POC-ize stuff with white people in it via AI. Doesn't matter what, redo the Friends on the couch. Make Captain America black (oh wai ), take it to stupid levels and give these fuckers the finger. Just don't make it racist stereotype replacement. If you can't tell what I mean by that, you probably shouldn't participate.
|
|
#
?
Feb 22, 2023 18:40
|
|
|
 female goblin casino dealer,
|
|
#
?
Feb 22, 2023 18:49
|
|
|
|
| # ? May 28, 2024 16:53 |
|
|
Fuzz posted:Let's do an AI Photoshop Phriday level idiocy event: I must admit I totally generated some what-if-I-was-black? versions of myself when I trained a model on myself, so I can do this immediately. My boyarsona and blacksona:  
|
|
#
?
Feb 22, 2023 18:53
|
|