|
I paid $70 for a logo of an octopus at one point For singular stuff I don't think I'll pay for that stuff going forward If there's been an uptick for custom human made art it's probably because a lot of people didn't know that cottage industry existed before, I'm not confident that increased sales trend will continue Am I gonna use MJ or SD to create the next "school of Athens"? No, but for generic yet specialized clip art it's difficult to beat especially for the price and turnaround time, and infinitely tweakable
|
# ? Apr 10, 2023 04:43 |
|
|
# ? May 28, 2024 00:51 |
|
update: no super crazy stream tonight, but I'm mutating Mario 3 right now https://www.twitch.tv/professor_dataviz Analytic Engine fucked around with this message at 05:40 on Apr 10, 2023 |
# ? Apr 10, 2023 04:45 |
|
Hadlock posted:If there's been an uptick for custom human made art it's probably because a lot of people didn't know that cottage industry existed before, I'm not confident that increased sales trend will continue Yeah, I don't think it's going to become a huge industry or anything, but there's always going to be some kind of mental work involved in translating thought to image (either through AI or traditional means) - it's not easy to write a prompt for exactly what you want, and even if we move away from text prompts there's always going to be some sort of work or creativity involved in getting a "perfect" image out of the AI be it by style-guiding with hand-drawn guides, clever prompting, personal pre-trained models, etc, which means there will always be a market for human artists willing to create the "perfect" image, because some people would rather pay for that than put all the time and effort in to learning the AI tools. The AI will be "good enough" for the vast majority of casual use. Using D&D character portraits as an example: the players using raw AI generated images are the same players who would have just found an image online before. The players who commissioned custom art of their character were never people who just "couldn't find a good enough image" or whatever - they were people who had grown attached to their characters and wanted to commemorate their accomplishments with a portrait drawn by an artist whose style they like. By commissioning an artist to draw it and having something that's truly "unique" (even if only in the way it was created), it becomes something personal and memorable, as opposed to just an image they slap on their character sheet. That market isn't going anywhere, but it's possible that it will shift towards artists who generate and tweak an AI image and then finish it by hand (and can therefore work faster at a lower per-image cost, which in turn makes it more accessible to potential customers, though who knows where that all balances out - art has never been great for business) deep dish peat moss fucked around with this message at 05:06 on Apr 10, 2023 |
# ? Apr 10, 2023 04:53 |
|
Analytic Engine posted:update: no super crazy stream tonight stream tonight, but I'm streaming V1 with Mario 3 right now
|
# ? Apr 10, 2023 05:00 |
|
Prolonged Panorama posted:Sure, but there is no "The Machines." Midjourney isn't "having a look" or "attempting" anything at all, that's anthropomorphizing software. Software isn't a moral agent, it just does what it's been programmed to do. Any anthropomorphizing was a figure of speech, but also a rather technically relevant one since neural nets truly are unique compared to most other software in that they are not programmed to do anything other than get really good at finding and recreating patterns. This is why, after all, they are called AI. Rather, they are tuned to, in this case, create depictions of plain language scene descriptions. Why I'm drawing this line is because my accusation of "tech racism" or "flesh-ism" or whatever becomes more formed; the Midjournies and Dall-E's of the world are black boxes to everyone, even their creators, and in that sense they "act" separately from them. The hard pill to swallow here is that visual art, and indeed all art generally, is nothing more than patterns on top of patterns ad infinitum, which like any other patterns can be learned, tweaked, and recreated endlessly. You can argue that the bots owe the quality of their output to the good patterns they learned from, free of charge, but to imply that all thus derived works are stolen is absurd.
|
# ? Apr 10, 2023 05:01 |
|
lol Internet bandwidth limit exceeded
|
# ? Apr 10, 2023 05:01 |
|
https://www.twitch.tv/videos/1789668583
|
# ? Apr 10, 2023 05:03 |
|
Also I haven't drawn much lately because I'm still waiting for a change in medication to even out but here's an example of the kind of thing I've been doing with AI recently: 1) AI-generated concept image: (This was MJv4) 2) Pixel art dialog portrait (unfinished), using the AI-generated image as a reference:
|
# ? Apr 10, 2023 05:17 |
|
deep dish peat moss posted:E: also, as an artist, people being able to AI generate a concept image of what they want rather than trying to explain it through words is a godsend
|
# ? Apr 10, 2023 06:35 |
|
deep dish peat moss posted:Also I haven't drawn much lately because I'm still waiting for a change in medication to even out but here's an example of the kind of thing I've been doing with AI recently: New Voynich manuscript looking good
|
# ? Apr 10, 2023 07:16 |
|
Pvt. Parts posted:Any anthropomorphizing was a figure of speech, but also a rather technically relevant one since neural nets truly are unique compared to most other software in that they are not programmed to do anything other than get really good at finding and recreating patterns. This is why, after all, they are called AI. Rather, they are tuned to, in this case, create depictions of plain language scene descriptions. Why I'm drawing this line is because my accusation of "tech racism" or "flesh-ism" or whatever becomes more formed; the Midjournies and Dall-E's of the world are black boxes to everyone, even their creators, and in that sense they "act" separately from them. The hard pill to swallow here is that visual art, and indeed all art generally, is nothing more than patterns on top of patterns ad infinitum, which like any other patterns can be learned, tweaked, and recreated endlessly. You can argue that the bots owe the quality of their output to the good patterns they learned from, free of charge, but to imply that all thus derived works are stolen is absurd. suttons bitter lesson is a special case of the bitterness peeps always have for behaviorism. rumelharts advisor was estes, estes advisor was skinner, so this is behaviorists telling you that you arent special with two layers of indirection (in the model sense, more like 25 layers) and being right, lol bob dobbs is dead fucked around with this message at 07:42 on Apr 10, 2023 |
# ? Apr 10, 2023 07:20 |
|
you got those extra layers of indirection so you gotta call it "cognitive" lol
|
# ? Apr 10, 2023 07:25 |
|
Fuschia tude posted:New Voynich manuscript looking good I tried generating a new Codex Seraphinianus back in the v4 days, good reminder to try again From v4:
|
# ? Apr 10, 2023 07:39 |
|
deep dish peat moss posted:I can safely say that after several years of intensely messing with AI generation tools, this is absolutely not the case. It is going to change the toolset they use, not obsolete them. Pretty much this https://youtu.be/xm7BwEsdVbQ Dude worked on the lion king, I think we can safely say he knows what he’s talking about when he says it’s another streamlining tool to get more work out there
|
# ? Apr 10, 2023 07:55 |
|
Thread needs more art. Some collages I made from my stuff.
|
# ? Apr 10, 2023 12:41 |
|
deep dish peat moss posted:Will it lead to smaller teams with more creative freedom? Yes. Should it require a team of hundreds of artists to tell a story in the first place? No. Will there be more creative content out there, both good and bad? Yes. Anything that reduces the costs means more stuff out there, and potentially better stuff if a bad thing doesn't need an extra $30mil to fix.
|
# ? Apr 10, 2023 12:45 |
|
More Corridor Digital talking about ethics in Image Generation https://www.youtube.com/watch?v=jG55yg3w0QU
|
# ? Apr 10, 2023 14:50 |
|
deep dish peat moss posted:I tried generating a new Codex Seraphinianus back in the v4 days, good reminder to try again What was the prompt for this? Getting dangerously close to Dr Seuss style art Edit: been thinking about making custom Dr Seuss books for my daughter and nephews
|
# ? Apr 10, 2023 18:24 |
|
Fuschia tude posted:New Voynich manuscript looking good "Donald Trump from the Voynich manuscript"
|
# ? Apr 10, 2023 19:45 |
|
epic bacon posted:"Donald Trump from the Voynich manuscript"
|
# ? Apr 10, 2023 21:39 |
|
That Bing Image creator is just showering me with credits. I've used like 100 credits today just screwing around making battlemechs and they gave me 100 more. Anyway, have some mechs: Also cockpits looking way nicer now: Nice Van My Man fucked around with this message at 03:58 on Apr 11, 2023 |
# ? Apr 11, 2023 03:51 |
|
What were the prompts for the cockpit views
|
# ? Apr 11, 2023 05:04 |
|
https://i.imgur.com/u0h87qr.mp4
|
# ? Apr 11, 2023 09:21 |
|
Nice Van My Man posted:That Bing Image creator is just showering me with credits. I've used like 100 credits today just screwing around making battlemechs and they gave me 100 more. Anyway, have some mechs: Hadlock posted:What were the prompts for the cockpit views
|
# ? Apr 11, 2023 09:30 |
|
I didn't get too creative with the cockpit, just "cockpit view from a helicopter fighting mechwarrior battlemechs." Helicopters make the best generic cockpit. Actually one thing I've noticed about Bing Image Creator that's the opposite of the older DallE is that I tend to get better results by keeping it short, whereas before I got better results by making it as verbose as possible. My results when I really try to describe parts in detail always end up really screwed up like that list 4 legged mech I posted. Also multiple subjects tend to flavor each other, so if its a helicopter cockpit but there are mechs in the scene it's naturally going to get more mech-ish.
|
# ? Apr 11, 2023 12:55 |
|
these katbernetes pods are NOT coming back up
|
# ? Apr 11, 2023 12:57 |
|
Cabbages and Kings posted:
Mark Zuckerpurrg
|
# ? Apr 11, 2023 13:07 |
|
I decided to try MidJourney to make some illustrations and stuff for my TTRPGs. Here's a batch of portraits of the characters in my V:tM game: Honestly pretty surprised that I managed to get this with just a fairly simple prompt. quote:<insert character description here>, portrait, in the style of Alphonse Mucha, <description of clothing here>, stained glass background, high detail, realism, 8k, lightning ::3 --ar 3:4
|
# ? Apr 11, 2023 13:10 |
|
Skios posted:I decided to try MidJourney to make some illustrations and stuff for my TTRPGs. Here's a batch of portraits of the characters in my V:tM game: Very nice. What’s the “lightning ::3” notation here? I’m not familiar with it
|
# ? Apr 11, 2023 14:18 |
|
Ben Nerevarine posted:Very nice. What’s the “lightning ::3” notation here? I’m not familiar with it It's MidJourney's syntax for multi-prompts with weighting: https://docs.midjourney.com/docs/multi-prompts I think in this example though, it's not doing anything. There's only one prompt, so there's nothing to normalize.
|
# ? Apr 11, 2023 14:33 |
|
Ima Computer posted:It's MidJourney's syntax for multi-prompts with weighting: https://docs.midjourney.com/docs/multi-prompts Thanks!
|
# ? Apr 11, 2023 14:36 |
|
Yeah, that was me just copying what a friend told me. Likewise with the 'lightning'. It was meant to be 'lighting', I typo'd, liked the result of the first one and decided to keep it.
|
# ? Apr 11, 2023 15:34 |
|
https://github.com/samb-t/infty-diffquote:We introduce ∞-Diff, a generative diffusion model which directly operates on infinite resolution data. By randomly sampling subsets of coordinates during training and learning to denoise the content at those coordinates, a continuous function is learned that allows sampling at arbitrary resolutions. In contrast to other recent infinite resolution generative models, our approach operates directly on the raw data, not requiring latent vector compression for context, using hypernetworks, nor relying on discrete components. As such, our approach achieves significantly higher sample quality, as evidenced by lower FID scores, as well as being able to effectively scale to much higher resolutions. They trained this on a 80GB A100; for fun I am attempting to see if I can train it on a 4090 using a smaller batch size.
|
# ? Apr 11, 2023 20:25 |
|
Hadlock posted:What was the prompt for this? Getting dangerously close to Dr Seuss style art The prompt for that was just "surreal diagram of a bizarre alien facility. vivid, psychedelic, colorful. codex seraphinianus." Here's a result of the same prompt in v5 - not a big fan of this image but you could probably get very Codex Seraphinianus/Voynich pages out of it with a little effort. As-is I like the v4 version more The mech cockpit images made me want to try the same thing in MJ - ever since v5 it has been great about taking camera position instructions Cockpit view from a towering battlemech in the middle of a cinematic war --ar 3:2
|
# ? Apr 11, 2023 22:12 |
|
I played a LANCER one-shot recently and this was my dude, codename Kerfuffle:
|
# ? Apr 11, 2023 22:32 |
|
Messing around with mech cockpit stuff some more in MJ5, it was easy enough to get the cockpit view composition in place but the content was trickier Cockpit view from a towering battlemech. Surreal xenoform landscape. Cinematic battle against an army of xenos. --ar 3:2 Cockpit view from a towering battlemech. Surreal xenoform landscape. Cinematic battle against an army of xenos. Cyan cockpit lights, ultra-futuristic hologram targeting sensor --ar 3:2 The dashboard gets better when I describe a cyan HUD instead of cyan cockpit lights, but it's still not generating an action scene Cockpit view from a battlemech. Mutant xeno-jungle. Battle against swarm of xeno-beasts. Cyan HUD, ultra-futuristic hologram targeting sensor --ar 3:2 Cockpit view from a hyperfuturistic battlemech. Mutant War. Dramatic war still. Swarm of mutant monsters. Crimson HUD, ultra-futuristic dash --ar 3:2 Getting away from the color filter effect from naming a specific color: Cockpit view from a hyperfuturistic battlemech. Mutant War. Dramatic war still. Swarm of mutant monsters. RGB HUD, ultra-futuristic instrumentation --ar 3:2 Cockpit view from a hyperfuturistic battlemech as it fires lasers at a swarm of aliens. Dramatic combat scene. Swarm of mutant monsters. RGB HUD, ultra-futuristic instrumentation --ar 3:2 It's still not firing lasers at aliens though Getting rid of the futuristic HUD because that was starting to look way to much like Tron Cockpit view from a heavy assault battlemech as it fires lasers at a swarm of aliens. Dramatic combat scene. Swarm of mutant monsters. Vintage HUD, hologram radar --ar 3:2 These last several images frankly suck, but the problem is that there's no environment described. Just adding a setting back in makes the images far better and finally gets some actual battle scenes: Cockpit view from a heavy assault battlemech as it fires lasers at a swarm of aliens. Dramatic combat scene in a blasted-out wasteland. Swarm of mutant monsters. Vintage HUD, hologram radar --ar 3:2
|
# ? Apr 11, 2023 22:36 |
|
Dragonball Z Harpo Marx
|
# ? Apr 11, 2023 22:40 |
|
Cabbages and Kings posted:for fun I am attempting to see if I can train it on a 4090 using a smaller batch size. note: I am now neck deep in having installed WLS2, installed CUDA 11.7 and conda properly and a bunch of other poo poo on a ubuntu VM under hypervisor, and am now, again, trying to get conda to process the requirements yaml without barfing the things I do for fun that I wouldn't do for half my usual hourly rate frantically staring at conda debug output. Invoke SAT faster, drat ye!!
|
# ? Apr 11, 2023 22:48 |
|
Cockpit view from Goku as he fights Vegeta. Cockpit view from Goku Battlemech as he fights Vegeta on Namek.
|
# ? Apr 11, 2023 22:52 |
|
|
# ? May 28, 2024 00:51 |
|
"President Donald J Goku" I have to admit this turned out way better than it had any right to. hydroceramics fucked around with this message at 23:16 on Apr 11, 2023 |
# ? Apr 11, 2023 23:13 |