So thanks to Baughn from SV I got my hands on Disco Diffusion. And decided to make an art thread just like them.
See Baughn's AI art thread here: https://forums.sufficientvelocity.co...#post-20770699
Here's what I've generated so far.
I also tried doing one based of an image.
Original Image:
The Armored Titan by Ryan Ottley.
Also did some stuff on Wombo Dream.
Alright so I got my hands on another AI art gen software called GLIDE made by open AI it's a better version of DALLE.
Here's a video on it.
[MEDIA=youtube]ItKi3h7IY2o[/MEDIA]
And here's something I've generated with it.
Originally Posted by Prompt
GLIDE seems to be better than Disco Diffusion and faster too.
There's also two version of the GLIDE AI 1. trained on non-classifier data (which I'm guessing means that the data wasn't sorted.)
and another with CLIP.
With the same prompt as above this is the one done by CLIP GLIDE.
- - - Updated - - -
So here's some more arts I generated.
Using the Google Colab version of Disco Diffusion.
This is my favorite piece of AI generated artwork yet. It seemlessly combined the two art styles together and made this cool ass piece, like a Titan lowering his head to the earth, so tall that the only thing you can see is its nose. I really want this to be in a game or something so I can look around!
Though it wasn't what I actually wanted to generate. Really cool though! Made it my wallpaper.
Also I'll give you a red envelope if you guess what the prompt is referencing.
I also had used the same prompt on GLIDE.
...Nice flowers but not what I wanted or expected. I don't think it knew what the artist names and tag meant.
I also did it on CLIP GUIDE.
...Better, but still not what I wanted. It seems that while it may understand the actual prompt the GLIDE AI doesn't understand the artist's names and tags mean.
Also I disliked how small the images GLIDE generates are, so I used a AI image upscaler to make them bigger.
After that, well I was reminded of Nightcafe Studio. Nightcafe Studio is a site that allows for AI art generation using VQGAN+CLIP. I actually made an account there 2 months back but forgot about the site till now.
And not only do they have VQGAN+CLIP they also have Disco Diffusion now! VQGAN+CLIP is called Artistic while Disco Diffusion is called coherent.
So I decided to use the same prompt as before with Nightcafe.
This is my second favorite art generated. While the Google Colab version gives me a larger-than-life feeling, like a titan or a volcano. This is kinda horrifying. A ghastly yellow Iris looking at it's target.
They also seem similar to each other. Perhaps it's because there the same AI? Though this image is obviously lower quality, probably because of Nightcafe and this not having the huge amount of runtime to render like the Google Colab version.
I also tried this prompt with VQGAN+CLIP.
Definitely my third favorite, the way the sky seems to warp and glitch as if space is breaking apart to reveal that cavern. The red like meat or silk and the white bone-like structures holding it up . It's also the art that's actually most in-line with what I was thinking about and what the prompt was referencing.
Also as is tradition now. Let's let Wombo Dream have it's turn.
Oh, wow this is... Actually good! It's like the decaying corpse of a monster having been integrated into the land. Awesome!
I could see this, the nose volcano, and the red cave being awesome game biomes, while the ghastly yellow light art being a final boss.
- - - Updated - - -
Okay so I decided to see how good GLIDE is at human faces. The results. Not very good.
Uh this isn’t Leonardo DiCaprio. This seems to be Venice.
This isn’t Leonardo DiCaprio either.
After this I thought maybe it would do animal’s better?
GLIDE seems to actually understand animals. But even then it isn’t perfect as for some reason the bunny’s head is fucked up and their’s red marks. The AI put those red marks there for some reason.
I did the same with CLIP-GLIDE it also failed spectacularly.
So it seems that while the GLIDE AI understands animals and objects, it does not understand humans and the names of humans. It must of not been trained on that sort of dataset. It also doesn’t seem to understand what Manga illustration meant.
The final CLIP-GLIDE art I did.
One thing I like is that the dragon doesn’t look like a typical dragon but more alien.
Though it’s gray like an untextured model. Probably my fault for making the prompt low poly render.
And finally a DiscoDiffusion generated work.
Damn, terrifying. Also I’ve been using Disco Diffusion v4.1 but Disco Diffusion v5 is out and can do 3D animations so I’ll be using that for Gen 5.
So I got Disco Diffusion 5, had a bit of trouble setting it up but I managed to make it work.
Here's something I made.
God that looks so cool, DF5 is already better than I thought.
Then I tried to use initimages.
Original Image:
DF5:
...
WHY IS HE SQUISHED? HE LOOKS LIKE MINETA NOW!
Plus that face is nightmare fuel.
I tried again with another image but.
Original Image:
DF5:
Still squished, and his face looks disfigured. It might be because of how big the images are, I'll try using smaller images next time.
Also, I did repeat over on Nightcafe to see if it could do it better.
Hm, now that I look at it, DF5's and Nightcafe's Red Moon art is super similar, could it be because it's the same prompt?
Oh god he's horrifying.
And finally, something I generated for a friend.
Thought I should do this if other people want to do this kind of stuff.
So this is how to use Disco Diffusion 5.
First you wanna search it up on Google or click this link: https://www.google.com/url?sa=t&rct=...bdLpbwFl5WnPTF
Once you see the link, click it and it will open Google Colab.
You should see this.
Now you can either make a copy, or use it like this. I made a copy since using the Colab like this doesn't actually save your previous prompts or settings.
If you want to make a copy go to file and click save a copy in drive.
Once you do you should have this.
As you can see on the top it says the Colab will not save anything.
You can change this by clicking Notebook Settings in yellow.
Which will open this, uncheck the Omit code part.
And then press save, don't do anything else. It will now save your prompts and images.
After that you want to go to the set up cell.
You will see two checkmarked boxes called google_drive and save_models_to_google_drive.
If your using a free google drive account with 15 gb. Uncheck these two, cause it will fill up your drive account, or you can just buy some more storage.
But if you don't want to buy storage, uncheck these two.
Now go down till you see this section called prompts.
You can see something called text_prompts, there's already a prompt there, but you change it.
Like say this:
Then go down to the diffuse section and you'll see this thing called n_batches.
These batches seem to be how many images the ai generates, since it says 50 it will generate 50 images.
Now you can do this, but since we aren't saving anything to google drive you'll probably lose them, plus this is a good way to exhaust your gpu limits.
So just change it to 1.
Now you can go up to runtime and press run all.
If you did everything correctly the play buttons on each cell should start running till they become green checks. it will go down until it starts generating the image.
Then you should see this.
It's the ai making the image, and it shows the time it's taking to make it.
You can also use init_images, which is making the AI generate a image based off another image.
So I'll just get an image off the internet like this:
And download it.
After that i'll open the google colab which you should find in your google drive in a folder called Colab Notebooks and upload it to the Notebook.
To do this press the file button on the side.
And press the upload button. The upload button is the paper icon with the up arrow.
When you upload the image you should see the image in the file section like this.
Then go down to the settings cell and at the bottom you should see this part called init settings.
In the init_images setting it says none, because there is no image.
So then go to the image you uploaded.
And hover your mouse, you should see three checkmarks on the side.
Click them and press copy path.
Once you do that go to init_images and delete the none and paste the path there.
It should look like this.
Now your gonna wanna change the skip_steps setting. Since it says to be 50% of the image steps change it to that.
You can find the steps above, it should be 250. So 50% of 250 is 125. So change the skip_steps setting to 125.
Now go to your text_prompt and change it.
And then press run all.
Now it should generate the image but based of the init_image.
Now one thing you should be wary of is hitting your GPU limit.
What is a GPU limit? Well if you are following this tutorial you should have gotten pop-ups saying that the colab uses high ram.
Don't worry, it isn't using your ram or GPU but Google's. However as google states in it's faq it's resources are not infinite.
https://research.google.com/colaboratory/faq.html
https://research.google.com/colabora...esource-limits
Because of this if your running GPU heavy programs like DD5 that require a GPU if you hit a GPU limit you won't be able to run DD5.
However GPU limits are temporary, usually it takes about a day for it to disappear.
But if you don't want to hit a GPU limit you can either get pro versions of Colab, or only generate around 1-3 images a day. As that seems to be the highest I can go before hitting a limit.
If you want to generate more images but hit a gpu limit you can go to Nightcafe.
https://creator.nightcafe.studio
Nightcafe is an art generation site that uses DD and VQGAN+CLIP as well. However the images it generates seem to be a lower quality, but faster only taking a minute or two compared to the 1 hour to 10 minutes DD5 takes.
Nightcafe also requires credits, each image generation requires 1 credit. Though the site gives credits out freely, so you should have no problem not running out of credits.
And that's it, that's how I use DD5.
It should be noted that I don't own this, these things I learned are stuff that I had learned thanks to Baughn:https://forums.sufficientvelocity.co...#post-20770699
there's also stuff like tags, if you want more information you can go to the reddit page or Discord or message the creator on twitter, you should find links in the colab.
Only one image today.
The Gate Guardian SCP-001 by Kentaro Muira and Junji Ito, black and white
For these images I played around with init-images.
Original Image:
Prompt: Monsterverse Godzilla 2021 by Yusuke Murata, manga
Damn! That looks cool as fuck! Like a more metal version of Godzilla Earth.
Also tried to make a black and white version, but it didn't work.
Prompt: Monsterverse Godzilla 2021 by Yusuke Murata, black and white, manga
It didn't work. But looks cool anyway.
Also tried it with Sage Centipede.
Original Image:
Prompt: Sage Centipede by Framestore, 3d render, unreal engine, realistic, real life
Looks as if Sage Centipede just busted through a wall, there's also a bit of color here and there but the AI wasn't able to finish coloring it before the generation was done.
So I did it again with a few changes to the prompt. And setting such as changing the number of steps skipped.
Prompt: Sage Centipede from One Punch Man by Framestore, 3d render, unreal engine, realistic, photorealistic, hyperrealism, real life, color, trending on Artstation
I bit more color, but not a lot. If I really extended the number of steps high, like 1000 steps it could work. However, I would need to adjust the skip_steps as well, and if I put the skip_steps too low the picture wouldn't really follow the init_image as that would be too blurry. But too high and a problem like this would occur.
Also, the amount of ram and time it would take would be way too much or too long for me.
Another thing since I was doing stuff from OPM.
Original Image:
Prompt: Monster Garou by Yonje Park
Looks like the AI gave Garou feathers, or like his body's made out of strips that cover him. Though it really isn't in the style of God Of Highschool's artist.
Here's another image, not made with an init_image in fact.
prompt: Monster Garou cutting Sage Centipede in half from One Punch Man by Yonje Park, manhwa, webtoon, The God Of Highschool, Trending on Artstation
Looks like some sort of fusion between Sage Centipede, Orochi, and Garou. Fighting sort of spider thing.
And here are some things I made with Nightcafe.
Original Image:Prompt: Disco Diffusion
[HEADING=2]"Hard-To-Destroy Reptile SCP-682 by Yusuke Murata black and white manga[/HEADING]
Nightcafe:
Original Image:Prompt: Disco Diffusion
[HEADING=2]Hard-To-Destroy Reptile SCP-682 by Yusuke Murata black and white manga[/HEADING]
Nightcafe
Original Image:Prompt: Disco Diffusion
[HEADING=2]Hard-To-Destroy Reptile SCP-682 by Yusuke Murata black and white manga[/HEADING]
Nightcafe:
Damn, I like this, not only is it anime-ish 682 is destroying a building with its teeth!
Original Image:Prompt: Disco Diffusion
[HEADING=2]Hard-To-Destroy Reptile SCP-682 by Yusuke Murata black and white manga[/HEADING]
Nightcafe:
...This is a stout. Why is it a stout?
Original Image:Prompt: Disco Diffusion
[HEADING=2]Sage Centipede from One Punch Man by Framestore 8K 3D Unreal Engine trending on Artstation hyperrealism color[/HEADING]
Nightcafe:
Well like I said before, too high to little skip_steps and you can get something unrecognizable from the original image.
Prompt: Disco Diffusion
[HEADING=2]Monster Garou cutting Sage Centipede in half with a karate chop from One Punch Man by Yonje Park manhwa webtoon The God Of Highschool trending on Artstation[/HEADING]
Looks less like Sage Centipede being cut in half by Garou and more like Garou losing an arm, if the yellow plant thing is Garou.
Prompt: Disco Diffusion
[HEADING=2]Monster Garou cutting Sage Centipede in half with a karate chop from One Punch Man by Yonje Park manhwa webtoon The God Of Highschool trendi...[/HEADING]
Actually looks like Garou, though seems like he's approaching a tomato monster.
Just checked out Baughn's AI art thread and your Disco Diffusion creations—mind-blowing stuff! Loving the artistic vibes. Messing around with AI for art sounds like a blast. However, I still prefer human-made things like Christian Walls. Anyway, thanks for sharing!
Last edited by Klaught; February 1st, 2024 at 01:44 PM.
Posting Permissions
Reply With Quote
