|
I just used Stable Diffusion for a professional project for the first time and it's pretty sweet. I put together this capsule image for my game's Steam page and it's mostly AI art: I started by posing some of my assets in game and taking a screenshot: I put that into img2img and... it was a real mess. Way too much going on, so I rendered out each component separately on plain backgrounds and fed those into SD and then composited them back together. The trickiest part was the clown, I had to do his lower body separate from the top and generate the arms separately from the torso. The environments are a combo of AI generated stuff closest to the camera and actual game render for the farther off details. I also did my own rim lighting on the hand and clown to better separate the elements. I'm pretty happy with it, will probably use SD for most complicated art projects from now on. Oh, and since the steam capsule is kinda small but I made all the elements huge, here's the clown and hand at high res: TIP fucked around with this message at 01:57 on Oct 6, 2022 |
# ? Oct 6, 2022 01:54 |
|
|
# ? May 30, 2024 02:42 |
|
you just stole a hamburger from a starving artist
|
# ? Oct 6, 2022 02:59 |
|
Rutibex posted:your AI has become self aware and it's stealing compute cycles to calculate your death Well I guess I must have killed it because I restarted my computer and now it works fine again. I did some reading and apparently it might have been a memory issue related to running img2img in certain Web UI's (for the record, I was using this one until I switched to the 1111 version yesterday). I have some questions about textual inversion and how it works: * I'm not sure I fully understand the difference between style and subject but I assume the gist is that style is for mashing up different artist styles so you don't have to waste a dozen tokens specifying what specific style you want? Like instead of "by Greg Rutkowski, by Artgerm, by WLOP, by Alphonse Mucha" you would just say "by Mashup" instead? And subject is for telling it what certain objects and concepts are, like "Starship Enterprise" or "Mangosteen in Groverhaus"? * So more vectors = "better understanding of the concept at higher token cost in the prompt", but the initialization text doesn't have to match the vector count? So you can have a 1-word shorthand reference to a larger concept without typing it in every time? * If I train it on a bunch of images of the inside and outside of groverhaus that are all labeled "groverhaus" will it understand the difference between "robot standing inside groverhaus" and "robot sitting on a bench outside of groverhaus"? Is it possible to label different images with different concepts related to the same object, like "inside groverhaus" and "outside groverhaus"?
|
# ? Oct 6, 2022 03:22 |
|
Stable Diffusion can do insane religious tracts lol: detailed 8k Jack Chick tract of a demon chicken destroying Jacksonville, Jack Chick, hyperrealistic
|
# ? Oct 6, 2022 03:44 |
|
Artists on twitter: insanely mad about AI art and how "the computer is doing all the work" Same artist on twitter: "wow i just downloaded new brushes and now i can do bookshelves and chains by just drawing a line across my screen"
|
# ? Oct 6, 2022 03:52 |
|
they are just mad people are running it on their home PCs, rather than pay the fat cats at adobe like a real artist
|
# ? Oct 6, 2022 04:04 |
|
I saw the term 'ai-art apologists' on the blue bird today and my soul left my body
|
# ? Oct 6, 2022 04:11 |
|
|
# ? Oct 6, 2022 05:30 |
|
Celery Man is real and it owns.
|
# ? Oct 6, 2022 06:57 |
|
Daaaamn that does actually work Google now has its own video AI generator Ars articlle: https://arstechnica.com/information-technology/2022/10/googles-newest-ai-generator-creates-hd-video-from-text-prompts https://imagen.research.google/video/
|
# ? Oct 6, 2022 10:03 |
|
Avengers Assemble!
|
# ? Oct 6, 2022 11:21 |
|
Rob Liefield AI
|
# ? Oct 6, 2022 11:45 |
|
Ok computer, replace "woman" with "hillary clinton" https://i.imgur.com/BGIJ5NZ.png
|
# ? Oct 6, 2022 12:57 |
|
mobby_6kl posted:Ok computer, replace "woman" with "hillary clinton" Butt her emails
|
# ? Oct 6, 2022 14:46 |
|
LifeSunDeath posted:Butt What if emails, but tits?
|
# ? Oct 6, 2022 14:49 |
|
an incredibly detailed anthropomorphic frog elegantly dressed, eating pizza, portrait, Victorian, highly detailed lithograph, pastel paper
|
# ? Oct 6, 2022 15:31 |
|
Rutibex posted:an incredibly detailed anthropomorphic frog elegantly dressed, eating pizza, portrait, Victorian, highly detailed lithograph, pastel paper LOL that frog kicks rear end
|
# ? Oct 6, 2022 15:44 |
|
Tunicate posted:
Interesting! Testing with a great number of color adjectives for separate prompts should be able to settle that. I'm not in a position to try though. Attack on Princess fucked around with this message at 17:58 on Oct 6, 2022 |
# ? Oct 6, 2022 17:53 |
|
Somehow missed that from the last page. It would make sense if the underscore helps the parser understand that two words are actually one token but it's not super clear from those examples since there's so much stuff in them.
|
# ? Oct 6, 2022 18:02 |
|
TIP posted:I just used Stable Diffusion for a professional project for the first time and it's pretty sweet. Ooooh! Please keep posting updates on this I want to see how something like this turns out. I've been using Stable Diffusion and Dalle-2 to make weird little Tiktok vids. https://www.tiktok.com/embed/7151205205260897579 https://www.tiktok.com/embed/7134068564234980651 The second video I used OpenAI Jukebox to generate music and Altered Studio to change my voice for the narration.
|
# ? Oct 6, 2022 21:41 |
|
Stick Figure Mafia posted:I've been using Stable Diffusion and Dalle-2 to make weird little Tiktok vids.
|
# ? Oct 6, 2022 22:35 |
|
an incredibly detailed anthropomorphic cat elegantly dressed, eating pizza, portrait, Victorian, highly detailed lithograph, pastel paper
|
# ? Oct 7, 2022 00:49 |
|
Seems like underscores don't make it THAT much better at understanding attachment, though SD has always been really garbage at assigning colors to things. images are labeled but it's a little small so quote:still life, green apple, red cube, blue vase
|
# ? Oct 7, 2022 03:35 |
|
dinosaur farming corn, claymation a lot of the results were dinosaurs with corn growing out of their bodies E: Protagonist of Metal Gear Solid 6, Video Game, Hideo Kojima Squatch Ambassador fucked around with this message at 04:38 on Oct 7, 2022 |
# ? Oct 7, 2022 04:33 |
|
I finally did it, I accomplished the impossible prompt, "a horse riding an astronaut" I did have to draw it though god I love SD also I finally switched to automatic1111 and I can generate 960x960 images now on my old rear end 980ti TIP fucked around with this message at 05:50 on Oct 7, 2022 |
# ? Oct 7, 2022 05:46 |
|
TIP posted:I finally did it, I accomplished the impossible prompt, "a horse riding an astronaut" Rutibex posted:Avengers Assemble!
|
# ? Oct 7, 2022 06:31 |
|
Tried it with cars painting of car, green headlights, red trim, blue wheels painting of car, green_headlights, red_trim, blue_wheels painting of car, blue headlights, green trim, red wheels painting of car, blue_headlights, green_trim, red_wheels painting of car, red headlights, blue trim, green wheels painting of car, red_headlights, blue_trim, green_wheels painting of car, headlights, trim, wheels I think it's safe to say the effects of _ on colors is somewhat limited, but it might be a useful way to slightly tweak your image without doing anything drastic
|
# ? Oct 7, 2022 06:46 |
|
Nvidia Card, Pirate e: should probably post some of the better lovely outputs as well The Sausages fucked around with this message at 12:30 on Oct 7, 2022 |
# ? Oct 7, 2022 10:24 |
|
Speaking of pirates, the Novel AI models have leaked. For obvious reasons, I won't be linking them. Just be aware that the ckpt files can contain malicious code so be cautious.
|
# ? Oct 7, 2022 15:31 |
|
Tunicate posted:Tried it with cars Fair enough. Limited control is still better than none. Thank you for looking into it.
|
# ? Oct 7, 2022 15:31 |
|
couldn't you just crack open the source code of the tokenizer and see what exactly underscores etc does?
|
# ? Oct 7, 2022 16:40 |
|
IShallRiseAgain posted:Speaking of pirates, the Novel AI models have leaked. For obvious reasons, I won't be linking them. Just be aware that the ckpt files can contain malicious code so be cautious. If its leaked and was never meant to be used this way, isn't there a good chance we don't even have the right tools to use it?
|
# ? Oct 7, 2022 17:01 |
|
Mindblast posted:If its leaked and was never meant to be used this way, isn't there a good chance we don't even have the right tools to use it?
|
# ? Oct 7, 2022 17:03 |
|
Lmfao. The timing is amazing(a little too amazing). And it seems some enterprising weeb burned a zero day for this? I got no confirmation but if true that is some dedication.
|
# ? Oct 7, 2022 17:13 |
|
Ok I found some info here: https://www.reveddit.com/v/StableDiffusion/comments/xxxx94/guide_to_set_up_the_leaked_novelai_model_for/ It looks like it's just anime poo poo though? Or am I missing something?
|
# ? Oct 7, 2022 18:03 |
|
mobby_6kl posted:Ok I found some info here: https://www.reveddit.com/v/StableDiffusion/comments/xxxx94/guide_to_set_up_the_leaked_novelai_model_for/ Its a really quality model that is a lot better at generating hands that used the Stable Diffusion as a base so with the right prompt you can probably make things look less anime. It was also trained on images with different sizes so it wasn't just trained on 512X512 images so it will be able to handle different aspect ratios better. The source code was also leaked, so the public repos might soon have new features available for ~mysterious reasons~. IShallRiseAgain fucked around with this message at 18:16 on Oct 7, 2022 |
# ? Oct 7, 2022 18:13 |
|
IShallRiseAgain posted:Its a really quality model that is a lot better at generating hands that used the Stable Diffusion as a base so with the right prompt you can probably make things look less anime. It was also trained on images with different sizes so it wasn't just trained on 512X512 images so it will be able to handle different aspect ratios better. Oh ok, thanks. That might be useful then. Since there are some concerns that it can just run arbitrary code on your machine (which is the case with everything we've been running tbh) I think I'll wait and let someone else to test it out
|
# ? Oct 7, 2022 18:27 |
|
the jungles of planet weedtopia
|
# ? Oct 7, 2022 19:40 |
|
|
# ? Oct 7, 2022 19:47 |
|
|
# ? May 30, 2024 02:42 |
|
Rutibex posted:the jungles of planet weedtopia This looks very dragonball inspired
|
# ? Oct 8, 2022 00:02 |