AI Art: It is criminal to not post your prompt

The Something Awful Forums > Main > General Bullshit > AI Art: It is criminal to not post your prompt

Chainclaw: Feb 14, 2009

KakerMix posted:

SPEAKING of Google, I've got into MusicLM and have been messing with it for the past few days. Every mp3 sample is 19 seconds and 24 kHz mono, though they've jumped from 64 kbps to 128 in the last few days. Even at this quality I'd be stoked to just have an infinite stream that generated music for me on the fly.

The system presents like this:

You input in the top left there, and you get two results in about 5 seconds, and it auto plays them one after the other. You can seek by clicking anywhere on the progress bar above the play/pause button, and can start and stop whenever you'd like or just jump right to the second one. The three vertical dots allow you to download the samples. No seeds, no instructions (besides the 'Try something like...') and really no way to know how you're supposed to use it. It is like any image generator you've used though, input text and press the button. It seems to know music instruments, types, generes, breakdowns and other music-things, much like Stable Diffusion knows cameras and angles and things. All of these are of course cherry picked, but I think this shows tremendous promise and very much feels like when I first started messing with MidJourney a year ago. Just needs more clarity, focus, tuning. It is censored so you can't just say 'The Beatles' or even do violence, but sort of can? Sometimes it trips up on 'funk' but other times it doesn't, and you can do deathmetal but not 'death'. Unfortunately it doesn't look like the prompt you use is saved to the mp3 in any form and the website itself also doesn't save anything once you close the window. I've started to save prompts in a .txt file to get around it, but I won't be able to tell you the exact prompt that was used to generate these. Some I remember what I was going for.

I'm dying for an AI model that generates midi and not pure audio. It would provide such a fast starting point for making music for small scale game development. I'd love to just put in something like "Hard driving shopkeeper music cigar chomping with an overdriven synth" and get a few midi tracks I could then bring into Ableton and polish out to something usable.

# ? May 23, 2023 18:27

Adbot: ADBOT LOVES YOU

# ? May 30, 2024 20:44

Cousin Todd: Jul 3, 2007; Grimey Drawer

We aren't far off I don't think.

Cousin Todd fucked around with this message at 18:42 on May 23, 2023

# ? May 23, 2023 18:40

pixaal: Jan 8, 2004; All ice cream is now for all beings, no matter how many legs.

Chainclaw posted:

I'm dying for an AI model that generates midi and not pure audio. It would provide such a fast starting point for making music for small scale game development. I'd love to just put in something like "Hard driving shopkeeper music cigar chomping with an overdriven synth" and get a few midi tracks I could then bring into Ableton and polish out to something usable.

Why not ask chat GPT to write the notes then feed that into something? Maybe some old software has a plain text save of sort that could be used if it's familiar with it?

# ? May 23, 2023 19:16

Chainclaw: Feb 14, 2009

You both had the same idea, and it's a decent one. It's actually the only thing I've actually used with ChatGPT so far. I wasn't too happy with the outcome, though.

I think the benefit of a more dedicated model is it would have a better understanding of how the various search terms relate to the note charts. I also figure there is just such a huge quantity of existing midi data out there that could be used. I think it's mostly the demand is too low for anyone to bother building it.

# ? May 23, 2023 19:27

Chronojam: Feb 20, 2006; This is me on vacation in Amsterdam
Never be afraid of being yourself!

pixaal posted:

Why not ask chat GPT to write the notes then feed that into something? Maybe some old software has a plain text save of sort that could be used if it's familiar with it?

Could've sworn somebody did that dozens of pages ago

# ? May 23, 2023 19:34

Mescal: Jul 23, 2005

Oh drat, I thought that music box was a continuous mix. I hope that's their goal.

# ? May 23, 2023 20:12

Mustang: Jun 18, 2006; “We don’t really know where this goes — and I’m not sure we really care.”

Tried making Robin Williams as Gandalf but got Tom Bombadil and Radagast instead.

# ? May 23, 2023 21:15

feedmyleg: Dec 25, 2004

Anyone else get the new Photoshop Beta going to test out the generative fill? ~~I just downloaded it but I can't find the tool itself where it shows here.~~

e: Apparently not only did I have to download the newest version, I also had to immediately Check for Updates despite it saying it was up to date, then install the new update. This is despite a tooltip immediately showing up calling out the feature after my initial download.

feedmyleg fucked around with this message at 22:04 on May 23, 2023

# ? May 23, 2023 21:58

pixaal: Jan 8, 2004; All ice cream is now for all beings, no matter how many legs.

this is an interesting challenge
(Robin Williams:1.2) in Gandalf cosplay costume, 4k

wait a minute how'd replace image size get checked under tiled oh no (I forget what model this was lost to sands of time) I think it was protogen using tile multi-diffusion with the default when you check overwrite image size (1024x1024) the many Robins are due to the tile and 1024 res that's not how that prompt was supposed to be run and it shows.

some model switching, this is the last image in a 6 batch on that seed
Robin Williams dressed as Gandalf , 4k, wizard, Bridge, action, casting spell, establishing shot, grey robe,
Negative prompt: captcha, text, sketch, scribble, error
Steps: 20, Sampler: DDIM, CFG scale: 6, Seed: 765253996, Size: 512x512, Model hash: 79939acf90, Model: verisimilitude_v2, Version: v1.2.1

Time taken: 17.44sTorch active/reserved: 3150/3770 MiB, Sys VRAM: 6001/8192 MiB (73.25%)

I think I'm gonna upscale him a few times this is a first run I'm not fully happy with but if I get a good pass it's easy to just iterate after getting the regions setup (I think, maybe, still new at this mostly why I'm interested in trying) Losing the robes as like waterfalls and landforms which would be cool if there was a portal I guess.

not what I was going for, if someone wants to keep taking this someplace it themselves this might form a rabbit hole by 4k. Part of me wants to see where this goes but that's hours and I'd rather explore a different latent space. (I do think a refined version of this could be very interesting)

pixaal fucked around with this message at 22:23 on May 23, 2023

# ? May 23, 2023 21:58

feedmyleg: Dec 25, 2004

Okay, 5 minutes playing with it in Photoshop, it absolutely hits the "good enough" barrier for everyday editing purposes. This is REALLY going to help my everyday workflow with pretty much every project.

# ? May 23, 2023 22:16

AARD VARKMAN: May 17, 1993

I'm making pokemon with a combo of ChatGPT and MJ + "niji" mode + "pokemon by Ken Sugimori"

Brineapple

Amerieagle

Goatloaf

# ? May 23, 2023 22:50

Dick Bastardly: Aug 22, 2012; Muttley is SKYNET!!!

AARD VARKMAN posted:

I'm making pokemon with a combo of ChatGPT and MJ + "niji" mode + "pokemon by Ken Sugimori"

Brineapple

Amerieagle

Goatloaf

these are so good

# ? May 23, 2023 22:51

pixaal: Jan 8, 2004; All ice cream is now for all beings, no matter how many legs.

AARD VARKMAN posted:

I'm making pokemon with a combo of ChatGPT and MJ + "niji" mode + "pokemon by Ken Sugimori"

Amerieagle

Had to look it up but I knew this seemed familiar
Braviary

I think you have a better design than Nintendo. Close enough they could make it one of the new regional variant things

# ? May 23, 2023 22:53

AARD VARKMAN: May 17, 1993

i tried giving him feet with adobe. I'm sorry for this

# ? May 23, 2023 22:55

pixaal: Jan 8, 2004; All ice cream is now for all beings, no matter how many legs.

just clone the arms

# ? May 23, 2023 23:02

Soulhunter: Dec 2, 2005

all hail denimthulhu

and disapproving dad in cargo jortsthulhu

Mescal posted:

Does anybody have an idea for placing characters in realistic (or at least in color) ancient Roman scenes? This is the first setting I've seen that it just completely chokes on. I'm using Bing.

Not much luck with Bing.
wearing athletic gear, ancient romans, in the year 100AD Rome, Mount Vesuvius erupts in the background

Better luck with Midjourney.

ancient roman cubicles in an office building

roman fashion

roman optimus prime

hipster using a laptop in a coffee shop in ancient rome

emperor wearing a VR helmet in the colosseum

e-sports in the colosseum

Soulhunter fucked around with this message at 01:19 on May 24, 2023

# ? May 24, 2023 00:49

Mescal: Jul 23, 2005

Does anybody have an idea for placing characters in realistic (or at least in color) ancient Roman scenes? This is the first setting I've seen that it just completely chokes on. I'm using Bing.

# ? May 24, 2023 00:55

Cousin Todd: Jul 3, 2007; Grimey Drawer

In case anyone wasn't aware, you can join the official stable diffusion discord and they have a free art generating bot. It's pretty decent.

https://discord.gg/stablediffusion

# ? May 24, 2023 03:15

feedmyleg: Dec 25, 2004

Mescal posted:

Does anybody have an idea for placing characters in realistic (or at least in color) ancient Roman scenes? This is the first setting I've seen that it just completely chokes on. I'm using Bing.

Not sure about Bing, but with MJ, I present a newly discovered Roman epic from the vault at Universal Studios:

All variations on "35mm still photograph from Spartacus (1960) of [x], Ancient Greece, trees, Ancient Roman forum, prehistory, togas, background image"

I'd imagine a similar approach would work.

# ? May 24, 2023 03:18

Cousin Todd: Jul 3, 2007; Grimey Drawer

Here's a video about a guy using gpt to create music

https://youtu.be/d_7EsKcn8nw

# ? May 24, 2023 04:05

Roman: Aug 8, 2002

There's definitely issues with making stuff attack other stuff. Many times I'll be like "monster attacking people" and they're all just hanging out together.

# ? May 24, 2023 04:22

Chainclaw: Feb 14, 2009

The Photoshop generative fill stuff is pretty impressive. It runs way better on our macbook pro than our PC, though.

edit:

I started with this image

and ended with this

edit:

A tiny bit more. He's vaping now, holding a vape pen, he got ketchup on his armor, he's got a cell phone clipped to his belt, he's got legs, and his skeleton has sunglasses.

Chainclaw fucked around with this message at 05:44 on May 24, 2023

# ? May 24, 2023 04:28

Bucnasti: Aug 14, 2012; I'll Fetch My Sarcasm Robes

feedmyleg posted:

Not sure about Bing, but with MJ, I present a newly discovered Roman epic from the vault at Universal Studios:

All variations on "35mm still photograph from Spartacus (1960) of [x], Ancient Greece, trees, Ancient Roman forum, prehistory, togas, background image"

I'd imagine a similar approach would work.

Some of those really capture the Ray Harryhausen look.

# ? May 24, 2023 06:47

KinkyJohn: Sep 19, 2002

Roman posted:

Here's a prompt for MJ I got from @nickfloats on Twitter that generates some cool horror stuff, even though it will often trip the auto mod and require a quick appeal: "50mm cinematic horror, darkcore [subject], Baltic violence tumblr --ar 16:9"

Computer says no. MJ moderation is flagging this prompt now

# ? May 24, 2023 07:37

Mola Yam: Jun 18, 2004; Kali Ma Shakti de!

I like the vintage cinemascope look. Was messing around with it and accidentally generated these cool helmet dudes.

edit: hell yeah, not at all what I was going for but I like it:

# ? May 24, 2023 08:08

Cousin Todd: Jul 3, 2007; Grimey Drawer

KinkyJohn posted:

Computer says no. MJ moderation is flagging this prompt now

Must be what you entered as the subject? I didn't hit even the first tier Automoderator putting turtle there.

# ? May 24, 2023 11:34

Kosmo Gallion: Sep 13, 2013

KinkyJohn posted:

Computer says no. MJ moderation is flagging this prompt now

A quick appeal usually gets this prompt accepted.

# ? May 24, 2023 12:32

hydroceramics: Jan 8, 2014

Roman posted:

There's definitely issues with making stuff attack other stuff. Many times I'll be like "monster attacking people" and they're all just hanging out together.

You have to get a little cute. For example, I find "playing tag" works better than "chasing" for some reason.

A t-rex attacking a school bus:

A t-rex hugging a school bus:

A t-rex playing tag with a school bus:

# ? May 24, 2023 12:35

Roman: Aug 8, 2002

PlaysGamesWrong posted:

Must be what you entered as the subject? I didn't hit even the first tier Automoderator putting turtle there.

yeah it still works for me. I did a bunch last night and actually only got it on turtle which I also tried, then it passed the appeal.

It's fun to put non-horror stuff as the subject in that. (50mm cinematic horror, darkcore [subject], Baltic violence tumblr --ar 16:9)

puppies

this was one from turtles?

capybaras

chickens

star wars

# ? May 24, 2023 12:39

Mescal: Jul 23, 2005

Does anybody know if you can sort Huggingface spaces by type, like text-to-image?

Here's one from recent that's just stablediffusion 2.1 with no options in the UI
https://huggingface.co/spaces/Dagfinn1962/stabilityai-stable-diffusion-2-1

This one seems cool with a number of options, but I'm not sure which options to use for general purpose. I don't know how to use the text to image, but it takes your image and makes a variation on it. It seems to want to make things more photorealistic in general. It runs alongside an auto-tagger here, which is fun. Any thoughts?
https://huggingface.co/spaces/heejun1213/zmlopsDiffusion

And Floyd If goon-recommended for text, has already been previously linked in this thread. Not a new one.
https://huggingface.co/spaces/DeepFloyd/IF

Thanks for the tips on ancient Rome! What model's good for a midjourney-like? Either free to use in Spaces/some website, or to run myself?

# ? May 24, 2023 21:09

KwegiboHB: Feb 2, 2004; nonconformist art brut
Negative prompt: amenable, compliant, docile, law-abiding, lawful, legal, legitimate, obedient, orderly, submissive, tractable
Steps: 32, Sampler: DPM++ 2M Karras, CFG scale: 11, Seed: 520244594, Size: 512x512, Model hash: 99fd5c4b6f, Model: seekArtMEGA_mega20

Mescal posted:

Thanks for the tips on ancient Rome! What model's good for a midjourney-like? Either free to use in Spaces/some website, or to run myself?

There's the old, (lol 6 months is forever ago) https://civitai.com/models/1123/midjourney-v4-paintart Midjourney v4 model for Stable Diffusion.
I'm surprised this is the only attempt at making a v5 model. https://civitai.com/models/50252/topnotch-artstyle-volume-1-18-midjourney-v5-images-trained-w-dreambooth-just-for-kicks
I've used the v4 model and it's good for what it is, I haven't tried the v5 though.

# ? May 24, 2023 21:32

KakerMix: Apr 8, 2004; 8.2 M.P.G.

AI art brought into the physical world and stuck on my car (these are real pictures just to be clear)

# ? May 24, 2023 22:34

Hadlock: Nov 9, 2004

"economics is a flat circle"

Somebody with a better imagination, help me out here

# ? May 24, 2023 22:58

Mescal: Jul 23, 2005

Check it out I made this cheeseburger with AI

And then I made this cheeseburger in real life!

These are really images btw

# ? May 25, 2023 00:22

Roman: Aug 8, 2002

50mm cinematic horror, darkcore batman, Baltic violence tumblr --ar 16:9 (also [joker])

That prompt also gives me some crazy disturbing non-Batman images. I'm using those for some other stuff. Like this one is actually one of the tamer ones:

# ? May 25, 2023 00:40

pixaal: Jan 8, 2004; All ice cream is now for all beings, no matter how many legs.

darkcore batman, grimdark, horror, 8k, cinematic quality
Negative prompt: captcha, text, sketch, scribble, error
Steps: 20, Sampler: DDIM, CFG scale: 16, Seed: 2396855738, Size: 896x504, Model hash: 79939acf90, Model: verisimilitude_v2, Version: v1.2.1, Tiled Diffusion: "{'Method': 'MultiDiffusion', 'Tile tile width': 96, 'Tile tile height': 96, 'Tile Overlap': 48, 'Tile batch size': 4}"
Time taken: 28.31sTorch active/reserved: 3797/4624 MiB, Sys VRAM: 7162/8192 MiB (87.43%)

could easily upscale this to full 1080p, most models don't like generating that large without repeating so I tend to do 16:9 smaller and then upscale.