User Name
Password
AppleNova Forums » General Discussion »

Stable Diffusion


Register Members List Calendar Search FAQ Posting Guidelines
Stable Diffusion
Page 2 of 2 Previous 1 [2]  Thread Tools
turtle
Lord of the Rant.
Formerly turtle2472
 
Join Date: Mar 2005
Location: Upstate South Carolina
 
2022-09-13, 13:26

Off-topic (click to toggle):
The captchas that REALLY annoy me are the ones where I have to identify the lions with their eyes open or something like that that isn't completely obvious.

Just let me proceed with my web thing please...

Louis L'Amour, “To make democracy work, we must be a notion of participants, not simply observers. One who does not vote has no right to complain.”
MineCraft? mc.applenova.com | Visit us! | Maybe someday I'll proof read, until then deal with it.
  quote
drewprops
Space Pirate
 
Join Date: May 2004
Location: Atlanta
 
2022-09-13, 21:45

Quote:
Originally Posted by Brad View Post
By the way, you know how captchas have been asking you to identify cars and crosswalks and traffic signals and trees over the past many years? Every time you answer one, you are providing labels to those images so that Google (and friends) can feed them into their own machine-learning models. They are building a massive library of billions of driving-related images to help power the software of self-driving cars. Google isn't relying solely on that torrent of data to drive a car, but they are using it to double-check their own custom AI software's behaviors.
Mind blown.
I just didn't think about it.
The older I get the more blown away I am by how many opportunities there are to make money by developing niche software.

...

Steve Jobs ate my cat's watermelon.
Captain Drew on Twitter
  quote
Brad
Selfish Heathen
 
Join Date: May 2004
Location: Zone of Pain
 
2022-11-24, 16:10

Here are some recent attempts of mine to Star Wars Clone Wars-ify the main Star Trek captains.

Unfortunately, through this exercise I've learned first-hand the lesson of how these models have (hopefully unintentional) biases due to the many thousands of barely-tagged images sourced from all across the internet. Black men appear to be very poorly represented in the main SD model, and I've really struggled to make a Sisko who looks anything like the man himself. I've run through lots of different inputs, and I've discarded tons of ugly and weird and malformed outputs trying to make him look good. Picard, by contrast, was pretty much perfect with the first few tries.

Let this be a reminder to everyone! Your output is only as good as the input when it comes to these highly-trained systems, and that is especially true for the original data used for the training. The data in these recent models are fine for making faces if you only want very famous people, white men, or sexy ladies.

Kirk


Picard


Janeway


Sisko

The quality of this board depends on the quality of the posts. The only way to guarantee thoughtful, informative discussion is to write thoughtful, informative posts. AppleNova is not a real-time chat forum. You have time to compose messages and edit them before and after posting.
  quote
drewprops
Space Pirate
 
Join Date: May 2004
Location: Atlanta
 
2022-11-24, 18:32

Representation comes slowly. These are terrific!


...
  quote
Brad
Selfish Heathen
 
Join Date: May 2004
Location: Zone of Pain
 
2022-11-27, 18:42

Because I clearly have nothing better to do on a holiday weekend, I spent all day yesterday trying to train a stable diffusion textual inversion for Star Trek's Worf and use it to generate images in his likeness. I think I, uhh, well I'm not exactly sure where things went wrong.

It could be that I over-trained my model on his distinctly Klingon forehead ridges. It could be that the base SD model has no flipping clue what to do with his forehead since the many thousands of photos of normal humans it was trained on have smooth heads. I tried an absurd number of prompts and generated over 600 images, and I never quite got what I was looking for. Maybe I'll experiment with this idea again some other holiday weekend. Meanwhile, please enjoy this giant collage of some of the "best" results I hand-picked from the lot.

By contrast, once again, Sir Patrick Stewart was super easy to synthesize. It only took a few minutes to crank out some original studio portrait masterpiece shots like these.

I've also been on a kick putting famous folks into colorful clothes and sweaters, or putting them on stage performing standup comedy. Did you know that Queen Elizabeth regularly performed at clubs since she was a young woman? Who's going to dispute this impeccable evidence? #totallynotfakenews

Note that the system often can't make sense of Trump's ridiculous comb-over. Every once in a while I got these hilarious interpretations of it. I swear it came up with those hairdos on its own. It also sometimes struggles to give Trump a not-melting/exploding/horrific face (none of this included here) presumably because so many photos over the years show him shouting angrily with his mouth wide open.

(I'm putting the images in the next post so this one doesn't get whacked out horizontal scrolling from their bigliness.)

The quality of this board depends on the quality of the posts. The only way to guarantee thoughtful, informative discussion is to write thoughtful, informative posts. AppleNova is not a real-time chat forum. You have time to compose messages and edit them before and after posting.
  quote
Brad
Selfish Heathen
 
Join Date: May 2004
Location: Zone of Pain
 
2022-11-27, 18:43

Scroll to the right! Lots of examples.








The quality of this board depends on the quality of the posts. The only way to guarantee thoughtful, informative discussion is to write thoughtful, informative posts. AppleNova is not a real-time chat forum. You have time to compose messages and edit them before and after posting.
  quote
drewprops
Space Pirate
 
Join Date: May 2004
Location: Atlanta
 
2022-11-27, 18:57

OMG I definitely want in on this technology. I have been playing with Midjourney and I want more control, but training the AI seems to be the key and it sounds like you need a blue million photo references?

EDIT: Worf of the Thousand Ridges

...
  quote
Brad
Selfish Heathen
 
Join Date: May 2004
Location: Zone of Pain
 
2022-11-27, 19:45

Quote:
Originally Posted by drewprops View Post
…but training the AI seems to be the key and it sounds like you need a blue million photo references?
If you wanted to train a whole model from scratch, then yeah, you need a millions pictures and a many thousands of dollars worth of hardware-time. Very few folks are creating whole models from scratch. What I'm trying to do is create a "textual inversion" which is an add-on file that influences the behavior of an existing model. My TI output is a tiny file that's only 4 KB compared to the multi-GB trained model files. From what I've read, you only need a dozen or so pictures to make a good TI, but like with everything else about these systems, the quality of its output is highly dependent on…

1. the quality of the pictures with some variety (different angles, backgrounds, clothes, facial expressions)
2. good textual descriptions associated with each picture
3. a lot of patience (and willingness to throw it all away and start over when it fails)

I was struggling to train/build the TI file on my own system. So, I used free credits on a Google Colab server to crunch the numbers. It took three and a half hours to process twelve 512x512 pictures, and I had to keep poking things on the page every few minutes to make sure it didn't think I was "idle" and shut itself down because their free credits are not intended for long-running, unattended jobs like this. Yes. It took 3.5 hours to make a 4 kilobyte file. The maths are weird.

In hindsight, I think I framed my images too closely on his face, and I didn't provide enough detailed descriptions for them. If I do this again (and I probably will), I'll try to run it all locally so I can let it crunch overnight worry-free.

Quote:
Originally Posted by drewprops View Post
EDIT: Worf of the Thousand Ridges
I think Chewbacca-Worf near the center might be my favorite unexpected reject.

The quality of this board depends on the quality of the posts. The only way to guarantee thoughtful, informative discussion is to write thoughtful, informative posts. AppleNova is not a real-time chat forum. You have time to compose messages and edit them before and after posting.
  quote
Brad
Selfish Heathen
 
Join Date: May 2004
Location: Zone of Pain
 
2022-11-27, 19:48

Also, since more and more apps are being released to extend the accessibility of SD tools to non-programmers, I'm moving this thread over to General Discussion…
  quote
drewprops
Space Pirate
 
Join Date: May 2004
Location: Atlanta
 
2022-11-27, 23:17

Thanks for the context.

I hope that everyone has seen the "Start Wars" and "Stop Wars" AI photo sets that display a wildly, strangely dreamlike vintage take on the Star Wars film production.

Versions of Chewbacca that bend your brain because they DO kind of look like Chewie but look NOTHING LIKE HIM.

R2-D2s of all styles, many with AT-AT like legs.

Steampunk-ish versions of other characters.

Every new image is wilder than the next.

INSANELY addictive to view.

The kind of thing that we value when they spring from the minds of artists and jarring to encounter as a result of "steered chaos".

I would love to take a bunch of photos of Captain Drew and push them into a place where I could see him in a wild array of situations, but I do not have the time in my life to learn this level of technology and my ego isn't as big as his, to chase that down.

So I did stir and stir on one an architectural model in Midjourney, marveling at how it came together and FINALLY found one that was gorgeous enough that I paid for an account so that I could justify using it as a cover for a book that I didn't even have an outline for.... but it would be so cool.

Investment and interest in this technological arena is ONLY GOING TO GET LARGER.

It is marvelously fascinating, and I am sad that I am not able to just stop everything and pour myself into it.

Of course, in 10 years the AIs won't need my help anymore





...

Steve Jobs ate my cat's watermelon.
Captain Drew on Twitter
  quote
Brad
Selfish Heathen
 
Join Date: May 2004
Location: Zone of Pain
 
2022-11-28, 00:46

I for one haven't seen those! Got some links handy for us rubes?
  quote
drewprops
Space Pirate
 
Join Date: May 2004
Location: Atlanta
 
2022-11-28, 01:59

Oh goodness.

I am nearly asleep and can only provide Facebook links right now.

Here is the alternate version of Labyrinth:

https://www.facebook.com/groups/offi...ibextid=S66gvF

Start War 1922
https://www.facebook.com/media/set/?...ibextid=ncKXMA

And Stop Wars
https://m.facebook.com/story.php?sto...ibextid=qC1gEa


...
  quote
709
¡Damned!
 
Join Date: May 2004
Location: Purgatory.
 
2022-11-28, 15:17

I see Mario's probably got a side gig going on over at Chippendales! Hilarity.

I really have to find time to play more with these tools – some of those Midjourney posts are absolutely mind-blowing. A Metropolis and Hellraiser mashup? Yes please.

So it goes.
  quote
Brad
Selfish Heathen
 
Join Date: May 2004
Location: Zone of Pain
 
2022-11-29, 13:58

I'm kinda sorta trying to make modern Disney/Dreamworks/Fox 3D animation style versions of the main Smash Bros roster. These are the cherry-picked best of the best, standing on the shoulders of dozens of garbage results. Some of these are a lot harder to coerce than you'd think. I couldn't get Samus to have cartoony/3D face despite tons of different prompt adjustments. I haven't included any Zelda or Peach attempts because their outputs so far looked nothing like them.

These are all just the plain outputs straight from the machine, too. If you put a little time into retouching and compositing in Photoshop, generated images like these would look even more amazing.

Mario or Luigi (Super Mario Bros)


Samus Aran (Super Metroid)


Pikachu (Pokemon)


Yoshi (Super Mario series)


Ryu (Street Fighter)


Three types of Link!

regular adult Link (Legend of Zelda)


super buff child Link (Legend of Zelda)


super buff adult Link (Legend of Zelda)

The quality of this board depends on the quality of the posts. The only way to guarantee thoughtful, informative discussion is to write thoughtful, informative posts. AppleNova is not a real-time chat forum. You have time to compose messages and edit them before and after posting.
  quote
drewprops
Space Pirate
 
Join Date: May 2004
Location: Atlanta
 
2022-12-05, 01:35

Maxed out my budget with Midjourney for the month

Elvis as Tutankhamun





...
  quote
Brad
Selfish Heathen
 
Join Date: May 2004
Location: Zone of Pain
 
2022-12-05, 11:52

Nice.

Is Midjourney still just accessible via a Discord bot? I think I joined their Discord a while back, but I didn't like having to fiddle with the bot in public.

I've joined several Discords trying to find better models and embeddings and to learn some tips and tricks along the way, and boy oh boy there is no shortage of NSFW content being created in there.

The quality of this board depends on the quality of the posts. The only way to guarantee thoughtful, informative discussion is to write thoughtful, informative posts. AppleNova is not a real-time chat forum. You have time to compose messages and edit them before and after posting.
  quote
drewprops
Space Pirate
 
Join Date: May 2004
Location: Atlanta
 
2022-12-05, 13:10

Yes, it's on Discord.

When you join you have no other option but to work in public rooms.

Once you reach the end of the your public trial you can opt to pay for more.

At that point you can converse directly with the Bot.

Where is the NSFW content being created? It isn't possible with Midjourney, unless I'm missing something!!!



...

Steve Jobs ate my cat's watermelon.
Captain Drew on Twitter
  quote
Brad
Selfish Heathen
 
Join Date: May 2004
Location: Zone of Pain
 
2022-12-05, 13:54

If you run stable diffusion on your local system, the sky's the limit. The official Stable Diffusion models 1.5 and older are completely unfiltered, having been trained on everything from the LAION 5B data set. The official Stable Diffusion 2.0 model is much better for general use, but the authors went to great lengths to remove a lot of material from its training set. Models using or based on 2.0 have far fewer famous people/celebrities and have been stripped of many "styles" based on specific artist names. I wouldn't be surprised if they also removed some NSFW content or keywords. It's no secret that they're trying to "legitimize" their models and are trying to position it in a way that appears to sidesteps some potential copyright concerns.

But the official SD models aren't the only models.

People are already training new models from scratch or have retrained the original SD models using additional content, almost all of which fall into that very grey copyright legality. I've seen models trained on specific niche content like Disney screenshots, Fortnite characters, and Pokemon pixel art to whole catalogs of images from ArtStation, DeviantArt, and the very NSFW Danbooru.

If you're running your own software, you can download and use and mix any of these models together for interesting results. There are over a thousand "diffuser" models currently hosted on Hugging Face (https://huggingface.co/models?library=diffusers), but a model is just a big file that can be hosted and shared anywhere else more privately. So, some forums and Discords privately distribute models, and some require paid memberships through services like Patreon.

It's safe to assume that Midjourney is also running its own custom model. It's hard to know if they based in on an older version of SD or trained it from scratch, but they clearly aren't interested in sharing their secret sauce recipe.

The quality of this board depends on the quality of the posts. The only way to guarantee thoughtful, informative discussion is to write thoughtful, informative posts. AppleNova is not a real-time chat forum. You have time to compose messages and edit them before and after posting.
  quote
Brad
Selfish Heathen
 
Join Date: May 2004
Location: Zone of Pain
 
2022-12-05, 13:57

Also, if you're curious about what's in LAION 5B, they have a demo site where you can search for the original images associated with the data. This can help creators with prompt engineering since some words (especially fictional universe or technical things) may have no images or may have little representation in the training data.

https://rom1504.github.io/clip-retrieval/

The quality of this board depends on the quality of the posts. The only way to guarantee thoughtful, informative discussion is to write thoughtful, informative posts. AppleNova is not a real-time chat forum. You have time to compose messages and edit them before and after posting.
  quote
drewprops
Space Pirate
 
Join Date: May 2004
Location: Atlanta
 
2022-12-08, 20:06

Renaissance Era Starfleet Juggling Academy

I'm getting better at this.

https://imgur.com/a/jkd7sq4

...
  quote
drewprops
Space Pirate
 
Join Date: May 2004
Location: Atlanta
 
2022-12-10, 06:53

I need to stop sharing my fun experiments on Facebook, as there are too many people who are objecting to the fundamental technology itself and are laying into me with great and profound judgment.

I'm trying to understand this technology, as it directly affects my livelihood. Might as well have some fun along the way.


...
  quote
drewprops
Space Pirate
 
Join Date: May 2004
Location: Atlanta
 
2022-12-16, 07:02

pirates drinking beer on surface of the moon renaissance festival with dark starry sky and planet earth looming in background, bright stars, steampunk lunar rover, cinematic, action, movement

...
  quote
Brad
Selfish Heathen
 
Join Date: May 2004
Location: Zone of Pain
 
2022-12-16, 14:21

Quote:
Originally Posted by drewprops View Post
Renaissance Era Starfleet Juggling Academy

I'm getting better at this.

https://imgur.com/a/jkd7sq4

...
I love it!

Quote:
Originally Posted by drewprops View Post
I need to stop sharing my fun experiments on Facebook, as there are too many people who are objecting to the fundamental technology itself and are laying into me with great and profound judgment.
Fortunately they're always welcome here on ye old AppleNova forums.

Quote:
Originally Posted by drewprops View Post
I'm trying to understand this technology, as it directly affects my livelihood. Might as well have some fun along the way.
And it's addictive as heck, isn't it? I forced myself step away from the tools for the last week because I was sinking way too much time into tweaking this or that, recreating the same prompt over again with subtle wording changes each time, comparing different sampling methods and step counts, scouring the web for alternate models, etc. And in the end, most of it was still junk not really worth sharing.

Quote:
Originally Posted by drewprops View Post
pirates drinking beer on surface of the moon renaissance festival with dark starry sky and planet earth looming in background, bright stars, steampunk lunar rover, cinematic, action, movement
Make it so.

The quality of this board depends on the quality of the posts. The only way to guarantee thoughtful, informative discussion is to write thoughtful, informative posts. AppleNova is not a real-time chat forum. You have time to compose messages and edit them before and after posting.
  quote
drewprops
Space Pirate
 
Join Date: May 2004
Location: Atlanta
 
2022-12-21, 00:16

Here's something that I wrote for my friends a few days ago about AI - I was all wound up and winging it, but that's where I was:

Quote:

I have definitely fallen on the 'dark side' of the AI debate, as I continue to probe the mechanics of the Midjourney engine instead of making the sign of the cross and swearing off AI-driven image production.
This isn't the first sweeping technological wave that I have encountered, and I know that it isn't going away. Shunning it won't stop it.
Tools WILL develop to allow ethical image scrapers to respect visual works that artists want to keep out of the spooky cauldrons of AI modeling subroutines, but those will take years to develop and standardize and implement, and they will only be respected by ethically-managed image scrapers - computer industry friends who are more learned that I say that unethical image scrapers are already being used to create imagery that is not safe to show in polite company.
But, by the time NOFOLLOW tags for image scraping ethical AIs are developed, it will be too late - AIs will have progressed to a stage where they have created an internal library of pieces and parts that make up the human world, and they will be able to generate imagery without relying on stealing your photos and putting photographers out of business.
Of course, the thing that most people don't understand is that a LOT of professional photographers were put out of business 10 years or more ago when iStockPhoto came along and empowered soccer moms to make big bucks in their spare time by uploading photos for use at far cheaper prices than the big stock photo houses were charging.
The democratization of content generation swept through every corridor of creativity, from music production to graphic design to animation to stock video to writing.
There are already AI generators that advertise right here on Facebook to provide affordable copy for people who would rather go to a machine than find a competent human.
I have not seen an outcry about this theft, yet.
Perhaps it is coming, but it won't stop the AIs.
The majority of complaints that I have seen online about AI art theft is centered around real world artists finding their styles purloined by the AIs. I would be mad as hell too, to see my unique style turned into something akin to an image filter - to see something that was uniquely ME debased into just another option - devalued so that I would have difficulty ever using it again to earn an income, living in fear that the next style I develop will fall prey to the same ravenous engines.
So, I really do GET IT when people are ready to set a torch to the AIs ranged in their obtuse and unimaginable hyperdimensional plane.
But stealing art isn't the only thing that these AIs can do. They can also model reality in unimagined ways, since they're just slamming shit together and holding it up to say "is this what you asked for?"
One of the avenues that I have been having fun with is exploring unusual situations at Renaissance festivals, since that community defines one of my peculiar affections.
Asking an AI to create a photo of a pirate themed rollercoaster at a Renaissance festival is not exactly a common photograph you might find on the internet. In fact, the only one I've ever seen is one that I photoshopped together myself, several years ago, for fun.
Yet, in less than 5 minutes the AI came bounding back to me with four perfectly remarkable options of a piratey-steampunky rollercoaster thundering past camera, filled with Renfest pirates.
Absolutely REMARKABLE to see the camera angles that the AI selects, the expressions it creates, the eleventy-thousand fingers it puts on the left hand of a person on the edge of frame.
The technology behind these AIs isn't just thievery, it is black magic. A few years ago computer scientists readily confessed that they didn't really understand everything that was happening under the hood when they set an AI out to solve a task.
I am not sure if anything has changed in that regard, but it is my understanding that there are an ENORMOUS number of computations and permutations that an AI will go through when assembling an image based on nothing more than words.
Unlike the Lensa app, which many people now have retrograde guilt for using (because of its thievish yoinking of artists' styles), the AIs that take written instructions and turn them into imagery are absolutely addictive.
Yes, they DID gather that imagery from somewhere, or at least, they gathered the modeled CONCEPT of the imagery from somewhere.
When they are set loose to create an image they are taking pieces and parts from a million places and fitting them together into something brand new.
There IS another conversation happening outside of the stern admonishment of Facebook users, to discuss the difference between machines learning to draw from existing examples and humans learning to draw from existing examples.
At what point is it acceptable?
Fifteen years or so back a friend made a living painting still lifes of wine bottles and cigars and pocket watches. I learned that there was a community of local painters who worked in that medium, because the subjects were popular among people who had money. Who was copying who, in the community? I think there was some competition and some debate.
That's common in the art world, it's common in the world of fashion too. Knock-offs. Tributes. Both are derivative.
And so the conversation of derivative art, the sourcing of material, all of that is something that interests me on top of the technology and the dizzying power to rush back to the AI to find out if the magic spell of words that you composed resulted in something magical, or something ghastly - both options show up with great frequency.
In a conversation in a Facebook group artists were discussing taking the unexpected results created by an AI and cleaning them up in Photoshop, and adding to them.
THIS is going to happen, whether you like it or not. There ARE going to be branded AIs creating images for you from their approved stock image libraries.
And those stock photographers who are going to get rich from this new technology?
Well, they probably aren't.
Remember, some of them are soccer moms who replaced professional photographers.
There will always be something there to replace us, unless we are prepared to move to the new technology.
And so I can't turn away like some of my friends.
This is going to affect what I do for a living, and it is going to affect my colleagues. If I can't be there to experiment and understand the new technology on the block, shame on me for being so parochial - and once people understand the technology better they may adjust their opinions of it at some point in the future.
Justice is always chasing after technology. Every invention that has come along has caused controversy, consternation, and chaos. Laws had to be crafted to address an issue that had never existed before the technology.
Industry adapts to new technology.
I never downloaded "free" digital music, but a lot of people traded in it wildly, justifying their reasons. That was always between them and their belief systems and the artists they stole from. It wasn't my job to judge them then.
Once the music labels and movie studios caught up to the technology they were able to capture it and make it easy enough for people to pay a fee to use a service instead of spending time downloading ill gotten booty.
Until then, it was the wild west. It never resulted in the destruction of the industries, only a change in how they worked. Many individual artists were absolutely destroyed by the uncontrolled marketplace, but the industry survived - especially those who adapted.
Earlier this week a friend provided the perfect analogy when he said "I tell ya, these drum machines are going to put drummers out of business!"
Pirates on a rollercoaster.

  quote
Posting Rules Navigation
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

vB code is On
Smilies are On
[IMG] code is On
HTML code is Off

Page 2 of 2 Previous 1 [2] 

Post Reply

Forum Jump
Thread Tools
Similar Threads
Thread Thread Starter Forum Replies Last Post
How stable are Hackintoshes? Luca Purchasing Advice 40 2009-07-25 16:13
Hackintosh users: How stable is your system? turtle Third-Party Products 11 2008-06-28 07:16
Don't tell me Leopard is stable! ezkcdude Apple Products 30 2008-02-20 17:37
Dynamic Assignation vs. Stable URL drewprops Programmer's Nook 9 2005-12-11 18:25


« Previous Thread | Next Thread »

All times are GMT -5. The time now is 05:16.


Powered by vBulletin®
Copyright ©2000 - 2023, Jelsoft Enterprises Ltd.
Copyright ©2004 - 2023, AppleNova