Register a SA Forums Account here!
JOINING THE SA FORUMS WILL REMOVE THIS BIG AD, THE ANNOYING UNDERLINED ADS, AND STUPID INTERSTITIAL ADS!!!

You can: log in, read the tech support FAQ, or request your lost password. This dumb message (and those ads) will appear on every screen until you register! Get rid of this crap by registering your own SA Forums Account and joining roughly 150,000 Goons, for the one-time price of $9.95! We charge money because it costs us money per month for bills, and since we don't believe in showing ads to our users, we try to make the money back through forum registrations.
 
  • Post
  • Reply
StarkRavingMad
Sep 27, 2001


Yams Fan

feedmyleg posted:

Have a link for that?

https://forums.somethingawful.com/showthread.php?threadid=4023088

It's kinda dead though, last post was a couple weeks ago

Adbot
ADBOT LOVES YOU

deep dish peat moss
Jul 27, 2006

Not AI art but today I discovered that chatGPT can straight-up write fully-functional plugins for RPGMaker, including comments explaining how to use it. This opens up a lot of possibilities :kiddo: Now to come up with something cool to do with it.


"Write a plug-in for RPGMaker MZ that tracks six custom attributes per actor"

code:
//=============================================================================
// Custom Attributes Plugin
// by ChatGPT
// Date: 2023-04-05
// Free for commercial and non-commercial use
//=============================================================================

/*:
 * @target MZ
 * @plugindesc A plugin that tracks six custom attributes per actor.
 *
 * @help This plugin allows you to track six custom attributes per actor.
 * You can access these attributes using script calls or by using them in
 * formulas in the database.
 *
 * Usage:
 * - Add the following notetag to the actor notebox:
 *   <customAttributes: attr1, attr2, attr3, attr4, attr5, attr6>
 *   Replace attr1-6 with the names of your custom attributes.
 *
 * - To access an actor's custom attribute, use the following script call:
 *   $gameActors.actor(actorId).customAttributes[attrName]
 *   Replace actorId with the ID of the actor and attrName with the name of the
 *   custom attribute you want to access.
 *
 * - To set an actor's custom attribute, use the following script call:
 *   $gameActors.actor(actorId).setCustomAttribute(attrName, value)
 *   Replace actorId with the ID of the actor, attrName with the name of the
 *   custom attribute you want to set, and value with the new value.
 *
 * Example:
 * - Add the following notetag to the actor notebox:
 *   <customAttributes: strength, agility, intelligence, wisdom, charisma, luck>
 *
 * - To access an actor's strength attribute, use the following script call:
 *   $gameActors.actor(1).customAttributes.strength
 *
 * - To set an actor's agility attribute to 20, use the following script call:
 *   $gameActors.actor(2).setCustomAttribute('agility', 20)
 */

(() => {
  const pluginName = 'CustomAttributes';

  // Parse the actor notetags and create custom attribute objects
  const _processActorNotetags = DataManager.processActorNotetags;
  DataManager.processActorNotetags = function (actor) {
    _processActorNotetags.call(this, actor);
    actor.customAttributes = {};
    const regex = /<customAttributes:\s*(.*?)>/g;
    let match = regex.exec(actor.note);
    if (match) {
      const attrs = match[1].split(',').map(attr => attr.trim());
      attrs.forEach(attr => {
        actor.customAttributes[attr] = 0;
      });
    }
  };

  // Add script calls to access and set custom attributes
  Game_Actor.prototype.customAttributes = function () {
    return this._customAttributes;
  };

  Game_Actor.prototype.setCustomAttribute = function (attrName, value) {
    this._customAttributes[attrName] = value;
  };

  // Initialize custom attributes when an actor is added to the party
  const _setupNewActor = Game_Party.prototype.setupNewActor;
  Game_Party.prototype.setupNewActor = function (actorId) {
    _setupNewActor.call(this, actorId);
    const actor = $gameActors.actor(actorId);
    Object.keys(actor.customAttributes).forEach(attrName => {
      actor.setCustomAttribute(attrName, actor.customAttributes[attrName]);
    });
  };

})();
e: For more complex requests, it will respond by saying:

"I can't write a plugin that will do this directly, but here are the steps you'd need to take to do that:
1. blahblahblah
2. blahblahblah"

But then you can start a new chat and just say "Create an RPGMaker MZ plugin following these instructions:" and paste its own instructions back to it, and it will do it.

I was able to get it to create a UI menu for attribute distribution using only three one-sentence prompts, that I can stick in a game as soon as I set up some art assets for it.

The only problem I haven't solved yet (haven't put much thought into it) is the token limit for longer plugins but that can probably be solved by building them incrementally across multiple chats


e2: When it cuts itself off you can just prompt it with:

"Complete this RPGMaker MZ Plugin for [plugin purpose]

[paste the code it wrote here]"

And it seems to work, at least so far, though I'm sure it has its limits.

deep dish peat moss fucked around with this message at 23:54 on Apr 5, 2023

Hadlock
Nov 9, 2004

deep dish peat moss posted:


e: For more complex requests, it will respond by saying:

"I can't write a plugin that will do this directly, but here are the steps you'd need to take to do that:
1. blahblahblah
2. blahblahblah"

But then you can start a new chat and just say "Create an RPGMaker MZ plugin following these instructions:" and paste its own instructions back to it, and it will do it.

lmao

"you're capable of doing anything, you just need to believe in yourself. Here are the steps on how to believe in yourself"

Soulhunter
Dec 2, 2005

Big ol' image dump time from Midjourney.

Chill animals:


Splatter paint animals:


Rad pirate parrot:


Muppets / Art crossover:


Scary Stories:



Moxie-style advert:


Mostly abstract art:




Magritte-like:


Modern marble statue:


forbidden love:


and a kickass turtle:

hydroceramics
Jan 8, 2014
Having messed around with chat gpt and bard, the writing portions bore me to death. But this coding aspect is so interesting. It makes me wish I had any reason to use it.

Roman
Aug 8, 2002

Random art dump with prompts by chatgpt. all prefaced with "a classic blockbuster sci-fi action adventure movie in the style of steven spielberg + robert zemeckis." and ending with "Photorealistic, movie still, cinematic, reimagined by industrial light and magic, sharp focus, high quality --ar 16:9"

These Guys and The Adventure That Is In Space For Some Reason



In the middle of a tense space battle, the ship's engine suddenly fails, leaving the crew stranded in the middle of deep space. With limited resources and time running out, they must find a way to repair the ship and get back home.



The crew discovers a mysterious object hidden deep within the planet's core. As they investigate, they realize that it holds the key to unlocking the secrets of the universe, and they must fight to keep it out of the hands of a powerful enemy who will stop at nothing to possess it.





The crew engages in a final, epic battle with the enemy, using all their skills and resources to try to save the universe from destruction. In the end, they are victorious, but at a great cost, and they return home as changed people, forever marked by their journey beyond the edge of the known universe.




Bonus: The Sheveirk. Is it a gun or is it a star destroyer? Why not both? Find out in The Sheveirk, oper loew!

Roman fucked around with this message at 03:55 on Apr 6, 2023

END CHEMTRAILS NOW
Apr 16, 2005

Pillbug
I wondered how well SD would do making electron microscope images. It did okay at that, nothing spectacular. Then I decided to include a dog.



Taking a peak through some salt crystals.

Tagichatn
Jun 7, 2009

I really shouldn't prompt when hungry.

LifeSunDeath
Jan 4, 2007

still gay rights and smoke weed every day

END CHEMTRAILS NOW posted:

I wondered how well SD would do making electron microscope images. It did okay at that, nothing spectacular. Then I decided to include a dog.



Taking a peak through some salt crystals.

Hadlock
Nov 9, 2004

Soulhunter posted:

Big ol' image dump time from Midjourney.

Magritte-like:



and a kickass turtle:


These are especially good

mightygerm
Jun 29, 2002



The GPT compression 'language' thing making the rounds lately is WILD.
https://twitter.com/gfodor/status/1643444605332099072
https://twitter.com/gfodor/status/1643415357615640577

Analytic Engine
May 18, 2009

not the analytical engine


shallowfake webcam is almost ready

busalover
Sep 12, 2020

deep dish peat moss posted:


e2: When it cuts itself off you can just prompt it with:

"Complete this RPGMaker MZ Plugin for [plugin purpose]

[paste the code it wrote here]"

And it seems to work, at least so far, though I'm sure it has its limits.

No, this doesn't always work, there's reddit threads on different ways to make it work that all fail. If the for-pay version had the ability to complete complex code, that would get me to sub.

Hadlock
Nov 9, 2004

busalover posted:

No, this doesn't always work, there's reddit threads on different ways to make it work that all fail. If the for-pay version had the ability to complete complex code, that would get me to sub.

:same:

It needs to be able to do about 800-1200 LOC and you'd have an absolute monster on your hands

Seems to crap out around 80 LOC which for python barely gets through the boilerplate

Tagichatn
Jun 7, 2009

Meta doing some AI stuff now too: https://ai.facebook.com/blog/segment-anything-foundation-model-image-segmentation/

Their segment anything model can segment objects in a photo automatically. I dunno what it's gonna be used for but they mentioned it could be useful in ar/vr to identify objects. It sounds like it could be useful for blind people too, identify stuff in a photo or even real life.

pixaal
Jan 8, 2004

All ice cream is now for all beings, no matter how many legs.


Tagichatn posted:

Meta doing some AI stuff now too: https://ai.facebook.com/blog/segment-anything-foundation-model-image-segmentation/

Their segment anything model can segment objects in a photo automatically. I dunno what it's gonna be used for but they mentioned it could be useful in ar/vr to identify objects. It sounds like it could be useful for blind people too, identify stuff in a photo or even real life.

If you had this working with Automatic 1111 it'd be really nice for in painting masks.

KillHour
Oct 28, 2007


Tagichatn posted:

Meta doing some AI stuff now too: https://ai.facebook.com/blog/segment-anything-foundation-model-image-segmentation/

Their segment anything model can segment objects in a photo automatically. I dunno what it's gonna be used for but they mentioned it could be useful in ar/vr to identify objects. It sounds like it could be useful for blind people too, identify stuff in a photo or even real life.

Greenscreen that works in any environment on a smartphone camera and isn't poo poo. Motion tracking of objects. Removing your ex from the only photo you have at Niagara Falls.

Porn, probably

Aramis
Sep 22, 2009



Tagichatn posted:

Meta doing some AI stuff now too: https://ai.facebook.com/blog/segment-anything-foundation-model-image-segmentation/

Their segment anything model can segment objects in a photo automatically. I dunno what it's gonna be used for but they mentioned it could be useful in ar/vr to identify objects. It sounds like it could be useful for blind people too, identify stuff in a photo or even real life.

Object segmentation is step 1 of a whole slew of computer vision tasks. Sorting ripe vegetables on an assembly line? Isolating the relevant parts of the image is kind of important.

Also, Meta has been one of the largest player in the AI space for quite a while now.

Aramis fucked around with this message at 17:31 on Apr 6, 2023

RPATDO_LAMD
Mar 22, 2013

🐘🪠🍆

Hadlock posted:

:same:

It needs to be able to do about 800-1200 LOC and you'd have an absolute monster on your hands

Seems to crap out around 80 LOC which for python barely gets through the boilerplate

this twitter thread is about the newer GPT-4 and not chatGPT but it is actually very poor at solving problems it hasn't 'seen' before:

https://twitter.com/cHHillee/status/1635790330854526981
this makes sense -- it is, after all, a text prediction model, not a reasoning model. it is not physically capable of thinking "why" something works the way it does, it just replicates combinations of words and meanings it's seen before

feedmyleg
Dec 25, 2004

RPATDO_LAMD posted:

after all, a text prediction model, not a reasoning model. it is not physically capable of thinking "why" something works the way it does, it just replicates combinations of words and meanings it's seen before

At this point, anyone who is (or acts) surprised about this is either an idiot or a bad actor.

RPATDO_LAMD
Mar 22, 2013

🐘🪠🍆
pop culture treatment of "AI" as always being some variant of a-person-in-the-computer definitely doesnt help

feedmyleg
Dec 25, 2004
And if someone is being their understanding of ChatGPT on their half-remembered memories of Terminator 2, I'm very comfortable labeling them as an idiot.

Doctor Zero
Sep 21, 2002

Would you like a jelly baby?
It's been in my pocket through 4 regenerations,
but it's still good.

D’ya like dags?

Doctor Zero fucked around with this message at 17:50 on Apr 6, 2023

Roman
Aug 8, 2002

Image dump from me trying to make fantasy movie images.
"Journey to the Midrealms." GET IT?















Roman fucked around with this message at 00:58 on Apr 7, 2023

Hadlock
Nov 9, 2004

RPATDO_LAMD posted:

this twitter thread is about the newer GPT-4 and not chatGPT but it is actually very poor at solving problems it hasn't 'seen' before:

this makes sense -- it is, after all, a text prediction model, not a reasoning model. it is not physically capable of thinking "why" something works the way it does, it just replicates combinations of words and meanings it's seen before

It's pretty good at drawing 3d cubes in python and rust

But yeah if it's not explicitly been mentioned in stack exchange it might struggle with it. The good news is, every conceivable real world problem has already been described, asked and answered on stack exchange so it's still highly usable

Soulhunter
Dec 2, 2005

Hadlock posted:

These are especially good

Midjourney does a really good job of copying Magritte style with a 'Son of Man' prompt. I like what it's producing with V5.

I used the /describe function to get this prompt out of the Son of Man painting:
"the idea of man of my thinking, in the style of rene magritte, applecore, figural distortion, green, harry watrous, iconic album covers, wandering eye"

when fed back into the system, that got me this:


then modified a few things to get the guitar / color shift:
"the idea of man of my thinking, in the style of rene magritte, guitarcore, figural distortion, red, harry watrous, iconic album covers, wandering eye"

that brought me to a couple variations that I iterated on from there, but the variation/re-rolls really didn't need much work to get decent results



Anyways, I also managed to get some fashionable LARPing gear out of Midjourney with "a knight wearing a suit of armor, wearing jean shorts over his armor, denim shorts, jorts, cut-off jean shorts". I'm not sure how necessary the repetition is in the prompt, but it's amusing that Midjourney made the logic jump from "knight" to "larper", presumably when it saw "jorts".



Tactical jorts, stained shirt, man-boob plate. It truly has it all.

Soulhunter fucked around with this message at 18:56 on Apr 6, 2023

pixaal
Jan 8, 2004

All ice cream is now for all beings, no matter how many legs.


Soulhunter posted:


Anyways, I also managed to get some fashionable LARPing gear out of Midjourney with "a knight wearing a suit of armor, wearing jean shorts over his armor, denim shorts, jorts, cut-off jean shorts". I'm not sure how necessary the repetition is in the prompt, but it's amusing that Midjourney made the logic jump from "knight" to "larper", presumably when it saw "jorts".



Tactical jorts, stained shirt, man-boob plate. It truly has it all.

the armor looks too new still. In Stable diffusion the phrase Homemade does a lot of lifting with costumes. Also worn and torn just thrown in at the end can help if it's really stuck because you are using an over trained person like Musk.

Oh for negatives I like new, pristine

SD phrasing but MJ should understand it just as well.

first SD 2.1 as you prompted:

adding home made made it really focus on the armor, I swapped the formatting to be less repetitive since SD really dislikes that and will hyper focus on it.
a knight wearing a home made suit of armor, wearing denim jean shorts over his armor

this is feeling like what you probably wanted but I'm actually interested in more cosplay in it.
a knight wearing a home made suit of armor, wearing denim jean shorts over his armor, scuffed, scarred, worn
captcha, font, pristine, new

Or maybe this is more what you wanted?

It's fun comparing the different model's output

e:
slammed the final positive prompt into Bing since it doesn't support negatives


crushing it Bing

pixaal fucked around with this message at 19:36 on Apr 6, 2023

Mordiceius
Nov 10, 2007

If you think calling me names is gonna get a rise out me, think again. I like my life as an idiot!

Roman posted:

Image dump from me trying to make fantasy movie images.
"Journey to Midrealm." GET IT?

Holy poo poo.

8one6
May 20, 2012

When in doubt, err on the side of Awesome!

Roman posted:

Image dump from me trying to make fantasy movie images.
"Journey to Midrealm." GET IT?

















Well now I'm just sad this movie doesn't actually exist.

Sedgr
Sep 16, 2007

Neat!





Fuzz
Jun 2, 2003

Avatar brought to you by the TG Sanity fund

8one6 posted:

Well now I'm just sad this movie doesn't actually exist.

So is David Harbor.

Soulhunter
Dec 2, 2005

pixaal posted:


crushing it Bing

Bing absolutely nailed it.

I've tried messing around in Stable Diffusion, but my images seem to come out very rough or stylistically all over the place. Takes a ton of iterations to get close to what I'm aiming for. Inpainting seems powerful but similarly requires a painful amount of iterations on parts of the photo piece-by-piece.

For example, it took me no time to get an image of a tank in Stable Diffusion, but ~25+ iterations on inpainting to add the brain to the top and make it actually look like a brain. Quality is still not quite where I'd want it to be, but the idea is conveyed. I've succeeded at creating a Think Tank.


Just out of curiosity I took this back to Midjourney and did a /describe on the SD image then fed the MJ prompts back into the loop:

an old tank with its brain in the desert, in the style of stylized realism, political imagery, interactive, innovating techniques, highly realistic, neurocore, deutscher werkbund


an image of a tank with a brain, in the style of realism with surrealistic elements, desertwave, innovating techniques, 8k, ivanovich pimenov, gary larson, matthias haker


an image of a tank with a brain in it, in the style of realistic landscapes, desertwave, photorealistic renderings, action-packed cartoons, humanistic empathy, stark contrast, precisionism influence


a tanks with the brain on it, in the style of realism with surrealistic elements, realistic and hyper-detailed renderings, desertwave, digitally enhanced, michal lisowski, cartoon realism, 32k uhd


Notably, all of the re-prompts erase the brain element completely when generating images, despite 'brain' being preserved in the prompts themselves.

pixaal
Jan 8, 2004

All ice cream is now for all beings, no matter how many legs.


Soulhunter posted:


I've tried messing around in Stable Diffusion, but my images seem to come out very rough or stylistically all over the place. Takes a ton of iterations to get close to what I'm aiming for. Inpainting seems powerful but similarly requires a painful amount of iterations on parts of the photo piece-by-piece.


everything I gave you is one shot from 2.1 base model I didn't do any img2img inpainting or 200 generations that's best of 4 and the rest in the sets are pretty similar.

SD is about how you prompt you don't describe things in simple English you kind of chop it up into concepts you want combined. You also really want a negative prompt anything at all in there makes it a ton better especially things that throw out complete garbage like captcha.

oh one very important thing with SD, the default sampler Euler is very old at this point make sure to switch that to DDIM or something from the DPM family, they should be included. UniPC is new and I should start messing with that.

pixaal fucked around with this message at 20:20 on Apr 6, 2023

Sedgr
Sep 16, 2007

Neat!

SD feels a bit like stacking colored glass lenses to make an image. You kind of venn diagram your way there by trying to manipulate natural language tokens and accounting for and taking advantage of training biases where you can.

Adjusting the weights, token order, etc feel like adjusting the focus on a telescope filled with probability.

busalover
Sep 12, 2020
Interesting reddit thread on creating consistent characters

https://old.reddit.com/r/StableDiffusion/comments/12djxvz/how_to_create_consistent_character_faces_without/

quote:

Stable Diffusion model already knows tons of different people. Why not cross them together? A1111 has two options for the prompt swapping:

[Keanu Reeves:Emma Watson:0.4]

this means that at 40 percent mark it will start generating Emma Watson instead of Keanu Reeves. This way you can cross two faces.

There is another option:

[Keanu Reeves|Emma Watson|Mike Tyson]

Split characters with a vertical line and they will be swapped every step.

Add details to the prompt, like eye color, hair, body type. And that's it.

Here is the prompt:

Close-up comic book illustration of a happy skinny [Meryl Streep|Cate Blanchett|Kate Winslet], 30 years old, with short blonde hair, wearing a red casual dress with long sleeves and v-neck, on a street of a small town, dramatic lighting, minimalistic, flat colors, washed colors, dithering, lineart

quote:

Another tip is to put them in the negative prompt. I think the general advice is to put the opposite gender into the negative prompt, but I don't think that really matters



Positive prompt: A woman walking on a road

negative prompt: Keanu Reeves, Mike Tyson



I've also seen people say they used made up names as it tends to draw from the same latent space



Positive prompt: A woman Joanna Camelsonzzz walking on a road

quote:

A fun little excursion into negative land: Put an artist name or theme that you like as a negative prompt and use no other meaningful prompts. Generate some images and describe the results that are common to those images in text. For example, I found the opposite of H. P. Lovecraft was something like "wedding photos, happy, affluent, champagne, sunny day, trimmed lawn, neat garden, blue skies, fluffy clouds"

Now use that text as a negative prompt that acts as a sort of style guide, all your images should come out with the same unique feel to them, and you can be very brief with the prompts on the positive side.

Doctor Zero
Sep 21, 2002

Would you like a jelly baby?
It's been in my pocket through 4 regenerations,
but it's still good.

Roman posted:

Image dump from me trying to make fantasy movie images.
"Journey to Midrealm." GET IT?

















Wanna binge the poo poo out of this you farging bastage. :argh:

Did you use actors as basis to get consistent characters? I think is see the dude from Stranger Things and the girl that plays Ciri in the Witcher.

Roman
Aug 8, 2002

Doctor Zero posted:

Wanna binge the poo poo out of this you farging bastage. :argh:
Did you use actors as basis to get consistent characters? I think is see the dude from Stranger Things and the girl that plays Ciri in the Witcher.
that is David Harbour. The girl is MJ's failed attempt at Chloe Grace Moretz.

Seems I just keep proving over and over that even with my very limited prompt skills, AI art is the best movie pitch assistant ever. I wonder if anyone trying to sell projects in Hollywood is actually using this stuff.

KakerMix
Apr 8, 2004

8.2 M.P.G.
:byetankie:

Roman posted:

that is David Harbour. The girl is MJ's failed attempt at Chloe Grace Moretz.

Seems I just keep proving over and over that even with my very limited prompt skills, AI art is the best movie pitch assistant ever. I wonder if anyone trying to sell projects in Hollywood is actually using this stuff.

It's soaked into every single pitch, story board, script rewrite and design studio at this very moment.
It is way too powerful of a tool, now imagine you are a giant media conglomerate that hates unions and needs to make number go up.
Anyone would be foolish to think otherwise.

Roman
Aug 8, 2002

KakerMix posted:

It's soaked into every single pitch, story board, script rewrite and design studio at this very moment.
It is way too powerful of a tool, now imagine you are a giant media conglomerate that hates unions and needs to make number go up.
Anyone would be foolish to think otherwise.
Yeah I figured. All I see on social media is supposedly creative people saying AI Bad, and no one is talking about how everyone that matters is already using it everywhere already.

Adbot
ADBOT LOVES YOU

Prolonged Panorama
Dec 21, 2007
Holy hookrat Sally smoking crack in the alley!



I guess if "everyone that matters" means suits, execs, money people? The people that couldn't meaningfully contribute artistically before, but now feel they can? The people that will drive the massive art department downsizing that everyone seems to agree is inevitable? I guess in that sense they're the ones that matter, as far as these tools go.

Sure, plenty of artists / traditional "creatives" are also using AI generators, but many also... aren't. Some because they don't want to, others because it is very little help in their actual job.

The AI takeover may be in our future, but on the visual side at least it's far from a fiat accompli.

  • 1
  • 2
  • 3
  • 4
  • 5
  • Post
  • Reply