|
I'm gonna start adding multiples of the clothing wildcards just to see what the AI thinks i'm asking for this ended up being "(Sweater Sundress:1.2)" (Robe Baseball Jersey:1.2)
|
# ? Dec 21, 2022 20:30 |
|
|
# ? May 23, 2024 14:20 |
|
midjourney anime mode trip report! it seems to do a very good job of individual characters, but if you mix two characters togeather (like a fight) than it gets very confused and blends them togeather. It also can't compose a fight scene. I give it a C- it feels like a custom anime enbedding of StableDiffusion single character: multi character: I'm curious what it will spit out if i ask it for "anything" or optical illusions. more experiments await!
|
# ? Dec 21, 2022 20:47 |
|
I think you just created a new fetish
|
# ? Dec 21, 2022 21:04 |
|
precision posted:I think you just created a new fetish Impossible legs is your fetish?
|
# ? Dec 21, 2022 21:22 |
|
Playing baseball in robes seems hard
|
# ? Dec 21, 2022 21:42 |
|
BARONS CYBER SKULL posted:Playing baseball in robes seems hard If you can play quidditch in robes, you can play baseball in robes.
|
# ? Dec 21, 2022 21:50 |
|
BARONS CYBER SKULL posted:Playing baseball in robes seems hard Only one way to find out...
|
# ? Dec 21, 2022 22:30 |
|
KakerMix posted:Not sure why unstable diffusion was needed, such a big target porn generator. There are plenty of options to do that now without that, and there will be more tomorrow. Send money in exchange for vague promises of anime tiddies, get ripped off. It's the Kickstarter way.
|
# ? Dec 21, 2022 22:36 |
|
Scams should only be run by humans 😤
|
# ? Dec 22, 2022 00:09 |
|
BARONS CYBER SKULL posted:Scams should only be run by humans 😤 I made an AI that scams Scam AIs (sells timeshares).
|
# ? Dec 22, 2022 00:26 |
|
I generated these in midjourney using "isometric cutaway" which is a fun phrase to play with. My friend and I were talking about a game we'd like to see made, a GTA-style game but set in a realistically-sized normal-USA area with a few accurately-sized small towns and a small city. I was talking about how it should be possible to use AI like this to generate Sims-style floorplans so that every building can have a unique and enterable interior without having to copy and paste stuff. "Grand theft auto screenshot of gun battle in small town street"
|
# ? Dec 22, 2022 01:50 |
|
feedmyleg posted:MidJourney. I don't know how to run that fancy poo poo on my computer. Though I did just read that I can do it since my mac is M1. Yeah, that's both too bad and awesome. Awesome because those are really good. Too bad because this is something still not really cracked on SD and I got excited for a second. :/ It's been a kind of community project with the dndai group trying to get a model working that can generate battle maps in SD. No ones really been able to crack it yet, not enough that it's all that usable. There are a couple models that are good enough at doing one specific style of battle map but we're still pretty far from being and to just say "battle map of a temple in the middle of a feywild forest" and getting something you can just toss on roll20 for a game. But I think it's possible, and that when someone cracks it this will actually be a pretty big deal.
|
# ? Dec 22, 2022 02:02 |
|
dang this is probably the best thing i've ever generated "realistic child in sandbox playing with real world war 2 d-day soldiers"
|
# ? Dec 22, 2022 02:13 |
|
Tumble posted:My friend and I were talking about a game we'd like to see made, a GTA-style game but set in a realistically-sized normal-USA area with a few accurately-sized small towns and a small city. I was talking about how it should be possible to use AI like this to generate Sims-style floorplans so that every building can have a unique and enterable interior without having to copy and paste stuff. If I was head of Google, I'd make a GTA-ish game that scrapes Google maps and stitches the content into one huge map.
|
# ? Dec 22, 2022 03:08 |
|
Gynovore posted:If I was head of Google, I'd make a GTA-ish game that scrapes Google maps and stitches the content into one huge map. I started doing this with a racing game and didn't get very far. Someone rode through a neighborhood with a gopro strapped to their bike and used photogrammetry to create an assetto corsa track. https://www.youtube.com/watch?v=4IHexdiiL3c
|
# ? Dec 22, 2022 03:16 |
|
IIRC many buildings have some kind of image rights associated with them which is why video games like Spider-Man which are set in real locations generally don't use real buildings.
|
# ? Dec 22, 2022 03:17 |
|
Gynovore posted:If I was head of Google, I'd make a GTA-ish game that scrapes Google maps and stitches the content into one huge map. with procedurally AI generated house interiors so i can rifle through my neighbors cabinets and judge their decorating
|
# ? Dec 22, 2022 03:21 |
|
KillHour posted:I started doing this with a racing game and didn't get very far. Someone rode through a neighborhood with a gopro strapped to their bike and used photogrammetry to create an assetto corsa track. this is really dope tho
|
# ? Dec 22, 2022 03:22 |
|
Sorry for the wrong window reply on the last page. I just noticed it. Wtf. Won't let it happen again Tumble posted:
All of this is so amazing. I generated a ton of gunfight scenes for the hotline Miami thing and I had some really good results when I added "muzzle flash" to the prompt Edit: forgot to ask, did you upscale any of those cutaway images? Would love to see them in more detail Crapple! fucked around with this message at 05:13 on Dec 22, 2022 |
# ? Dec 22, 2022 04:56 |
|
Crapple! posted:All of this is so amazing. I generated a ton of gunfight scenes for the hotline Miami thing and I had some really good results when I added "muzzle flash" to the prompt And you're not sharing them? You monster.
|
# ? Dec 22, 2022 06:05 |
|
KillHour posted:And you're not sharing them? You monster. yeah it's a bit weird that none of them made it into the bunch that I shared. This was one of the cooler images I generated, but I just didn't think it quite fit in with the other pictures. Plus the 2 guys on the right appear to be having mechanical issues with their weapons
|
# ? Dec 22, 2022 06:46 |
|
I tried Stable Diffusion with "Talk poo poo post your fit."
|
# ? Dec 22, 2022 07:37 |
|
CRISIS COMES TO EARTH IN 100 DAYS. OUR ONLY HOPE LIES WITH FOUR ANCIENT WARRIORS AND THE ADMIRALS WHO LOVE THEM. CHOOSE YOUR BATTLE CAT WISELY. realistic historic world war 2 pearl harbor with a giant 300 foot tall cat rising from the ocean Crapple! posted:Sorry for the wrong window reply on the last page. I just noticed it. Wtf. Won't let it happen again I did not, but I am still mucking about trying to get it to use actual furniture or at least closer to it. When you zoom in it gets weird because it's obviously not actually making floorplans, it's stitching other bits of random isometric stuff it's seen and so stairs melt together, couches aren't couches, and so on and so on.
|
# ? Dec 22, 2022 08:07 |
|
this is incredibly fun to mess around with like that feeling putting in a precise prompt and getting exactly what you were thinking of is beyond satisfying and probably addictive now that i think about it. dopamine on demand
|
# ? Dec 22, 2022 08:35 |
|
it won't let me upload multiple in one post:(
|
# ? Dec 22, 2022 08:36 |
|
yea gimme some of that warhol warehouses
|
# ? Dec 22, 2022 08:37 |
|
I'll stop multi posting but this one's cool, i lost the prompt but it was something really close to, in dall-e 2: pen and marker illustration, vapor wave styling, late 80s japanese pop-up headlights sports car with street racing body modifications, livery has miami neon sunset color pallette, vivid contrast on car livery, thick outlines on car, minimalist shading, coloring on car overlaps lines, (i might've just said color it kinda messy) realistic lighting and reflections, headlights are bright, minimalist color use (i forget if i specified car, bg, or both), negative space and color saturation emphasized in background, high detail, entire subject visible in frame
|
# ? Dec 22, 2022 09:04 |
|
https://tactileturn.com/products/nexus I'm not mad about AI art but I'm livid about these loving dumb pens quote:These three digital images were generated in seconds by Artificial Intelligence through basic word prompts. Pretty crazy, right? With the Nexus, we wanted to highlight the ever-blurring line between humans and machines.
|
# ? Dec 22, 2022 09:44 |
|
they've been doing these exclusive limited-time releases of $150 pens for 2 years and they're all just palette swaps for 1 or 2 models of their normal pens they're also basically just a fancy titanium case around a pilot g2, of which you can buy a 5-pack for $6 on amazon right now
|
# ? Dec 22, 2022 09:58 |
|
BARONS CYBER SKULL posted:(Robe Baseball Jersey:1.2) BARONS CYBER SKULL posted:Playing baseball in robes seems hard KillHour posted:Impossible legs is your fetish? strange feelings re Daisy fucked around with this message at 10:03 on Dec 22, 2022 |
# ? Dec 22, 2022 10:00 |
|
Cabbages and Kings posted:I'm livid about these loving dumb pens the penis mightier than the sword
|
# ? Dec 22, 2022 10:11 |
|
well baseball is an american sport the gun is kinda assumed
|
# ? Dec 22, 2022 10:13 |
|
"tuxedo cat, crayon, line art"
|
# ? Dec 22, 2022 10:36 |
|
https://vimeo.com/783156438 And I finished it, the final version of this project. Things I learned doing it. Above all else, it's very hard to steer SD in the direction you want with img2img vs txt2img in a few ways. There were times I had to disable color coherence in Deforum in between sections because the last frame of the previous section ended up with a particular color dominating it that wouldn't work for the next section, or wasn't varied enough, but img2img really wants to keep the colors from the source image. Even with prompts that are just colors with colors I don't want as negative prompts, it just takes a lot to get it to actually change. It's a lot easier to get variety from images of women than it is for men. Especially when you want to keep it consistent over a bunch of frames. I'm just guessing there's more artwork/photos of women than men in the training data. You can get any style of any woman expressing any emotion, but not always for men. Also, a lot of stuff that can generate great images from noise just doesn't work how you would expect from img2img, at least with none of my starting frames. I guess that makes sense, that's how it's all designed to work in the first place. You kind of have to plan things out more too, because even if you can get good stuff out of it you're just necessarily more limited. Even with low strength, high noise, there are elements of the first frame the AI will just refuse to get rid of even over a hundred frames. I don't think I really figured out any kind of best practice. I developed a work flow that generally made things fast but still, sometimes I'd have 14 seconds of video done in 10 minutes and sometimes it'd take 3 hours. There's a lot to learning what you can put in prompts that will show up in the same way even if different prompts for keeping some consistency. The magic word I got from the thread "bioluminescent" would often show up as a glowing orb in a bunch of different prompts and would often take over the faces of the anime style characters, so I used that to maintain some kind of style throughout it. I think I'm going to keep up with this, I felt like the end result isn't just "something a human could have done but faster." A human isn't gonna draw 9000 wildly different frames in a variety of different styles to make a 5 minute video. So, it's a (as far as I know) somewhat entirely new style to animation. I'll probably try to do things that are more grounded in the future, that keep a more consistent style, but this was just going all over the place seeing what I could make it do. I just wish SD had a better understanding of verbs, being able to give it more complex directions would make things a lot easier. Same with being able to understand prompts with multiple subjects but where some other parts of the prompt apply to one subject but not the other. I had a lot of almost cool ballet scenes ruined by the man in the prompt wearing a tutu.
|
# ? Dec 22, 2022 12:05 |
|
BrainDance posted:https://vimeo.com/783156438 this rules, thanks for the breakdown
|
# ? Dec 22, 2022 12:12 |
BrainDance posted:https://vimeo.com/783156438 Good stuff
|
|
# ? Dec 22, 2022 12:32 |
|
Changing styles are done for collaborations https://www.youtube.com/watch?v=jjU_D-j7Qso There's still the AI transitions instead of clean scene cuts yours is different but wildly different styles In a single video can be done by humans. pixaal fucked around with this message at 13:06 on Dec 22, 2022 |
# ? Dec 22, 2022 13:02 |
|
pixaal posted:Changing styles are done for collaborations Yeah, I can think of some examples from even regular animation. I was more referring to the change in between individual frames like these 3 frames Where, because everything is always radically changing every frame you get an effect like the animation is crawling or something, I don't know the word for it, and that that's how it actually gets most of its animation, and that it's the whole scene changing as much as it does each frame instead of individual parts of it. In one way it's a flaw of the AI, not being able to keep much consistency even between three neighboring frames but it's its own style too and I think it's kinda cool if you just go along with it. I'm sure its been done by hand too, but I'd think it'd be pretty hard for a person to keep up with over a long video, especially if someone leaned into it hard and chose to go with low strength/high noise or something to intentionally get that style.
|
# ? Dec 22, 2022 13:39 |
|
There's certainly a ton more going on with AI, it just seemed like you thought the idea of wildly different styles hadn't been done. If you mean the morphing yeah it's not that easy an effect to get manually.
|
# ? Dec 22, 2022 14:17 |
|
|
# ? May 23, 2024 14:20 |
|
Ugh I wanted a fat Seagal asleep on toilet with a big burger next to him, but the AI has him kinda turning into food and maybe having eaten himself?
|
# ? Dec 22, 2022 15:53 |