Mastodon Politics, Power, and Science

Monday, April 3, 2023

A discussion about what ai image generation needs to be really useful. How to have ai generate text prompts from a story to input to an image generator.

 Him 

I like those two you did, that's awesome haha

Me

thanks! I am looking at how to create a text prompt my looking at examples.  It is almost like a magic incantation. LOL

Me

I found some negative prompts to not have detached deformed  limbs and extra fingers strewn around the picture. L

Me

I think it is funny that you have to specifically tell it not to do that.

Him

Haha yeah. I want to try asking ChatGPT to generate prompts for me

Me

I wonder if you have a picture described as a prompt with a specific seed if that could be considered a form of compression. L

oh, that could be cool.  Given this story, generate a series of prompts to feed into a specific ai image generator to illustrate the story.   Or to generate the panels for a comic book, putting in the text bubbles with the dialog.

It would be really cool it you could use the description of a character to create a consistent character across panels.

like a #define Harry_Potter <text prompt character description>

Me

Can you imagine downloading a comic book  from online as a 100k text file and it is generated  just in time as you turn the page.

Him

Woh that is an awesome idea. I know with Midjourney you can provide the seed (1-1000) so that it generates the same image every time

All of this is going to be come reality in the next 10 years or so as people figure out how to get AI to do things like this.

Imagine giving an AI a prompt for a story and it churns out a 300 page illustrated novel. "Like lord of the Rings but more battles and give it an unexpected twist ending"

Then a hardback gold foil "collectors edition" book shows up on your doorstep the next morning.

Hooking AI text and image models with KDP book printing and Amazon same day delivery.

Oh, and alsp make this one into a movie please.

Lol, the future is going to be wild.

himstory1.png

story1.png

I asked ChatGPT to give me a story, I fed each paragraph into Midjourney, and then I put it together in photoshop

Him

Somehow the input into each image needs to be all the previous images so that the characters are the same.

There definitely needs to be some logic that says, "this character" in "this scene"

I also can't get chat gpt to output real stories. It's always just "once upon a time" summaries if stories.

Him

Explicitly asking for character names and dialog helped though

An AI based “comic book maker” that takes prompts, writes a story, generates images into predesignated comic book panels, needs to exist

$$$ right there haha

Him

In time I'm sure that they will nail these edge cases down. Hands, weapons, fingers especially need work.

Just need more specialized training data

The fact that it does the big things super good, like "mood" "feeling" "emotion" but that it struggles on specific details is the opposite of what I would have expected.

But also a sign that there is no fundamental limit with the current architecture. If it can figure out the difference between a "somber" picture and a "epic" picture, then they will solve the 6 fingered hands lol

Me 

I actually get better results just by giving it more general  prompts  turning up the creativityand then having it create 40 images overnight.

Him

Ah okay thats a good point. What matters is that there's one good images, doesn't matter too much even if 39 of the 40 are bad, as long as you get the one great one.

That only matters if you are the human spending your own time painting all those images haha

Me 

I think if they create a poser model and then pose that in a scene with human anatomical limits and use that as a seed for the rest of the image that would help.


Tuesday, March 28, 2023

Storm damage.

 Ladder got blown over by an 80 mph gust of wind. It  broke the arm of my porch swing. I glued the arm back together to use as a template for a new arm.



Spring is sprunging.

 Got a lot of plants started and planted already.







DiffusionBee ai image generator

 This program was an effortless install on my macbook air. It downloaded the models itself and let me add in another one.  It only took about 20 minutes to generate an oil painting of a flower. 



Wednesday, March 22, 2023

Running AI locally on your own computer

 Was looking into what AI program I could run locally on my computer. 

I found https://github.com/openai/whisper 

Followed the directions to install and run it. And it blew me away. It can translate just about any language into time stamped English text. Even if there is background noise. Even if the language is being sung.

I easily turned a few music videos into karaoke songs by creating srt files that VLC displayed over the video. Tough songs that I could barely understand. 

I had to give it a --patience .85 setting because it would get stuck on really challenging parts of the music. Better to get 99% right quickly than 100% sometime tomorrow.

I also was able to record me speaking Spanish and have it create a sub titles in English.

Monday, March 20, 2023

Still living the life soltastic.

 I got a 12 x 16 shed  that had a tiny window in one wall and was bare stud walls. I insulated, wired, sealed it, drywalled it, and put real wood paneling  on the ceiling.  Then I added shelves clear around the whole place. Built a big workbench. And put a light strip up powered from solar panels. Once I get a little more organized it will be a perfect woodworking shop.




This is what it originally looked like, after I added the two reused rv windows.


Variables have type. They should have units too.

A conversation on computer science I had today with a friend.


 OK, one of the things that always bothered me with computer languages is that data has a type, but not a unit.


I completely agree

I have thought about that as well


I would love to see a computer language that could automatically convert miles to km if used miles in the calculation, but the result was in km


It would catch errors much earlier in the process


at least throw an error

orange = 5 apples....   unit error, compilation halted.

One of the space craft nasa sent up crashed at full speed into the martian desert because one programmer used miles and another used km.


Wow, that's crazy. I'm surprised they didn't catch the trajectory issue earlier on.

But yes. 5 = 5 is rarely what is being compared

Only when you are comparing truly unitless numbers, which is rare.

It's almost always 5 of something = 5 of something.


Yeah, I have only found a few formulas where all the units cancel out and give you a number

I think ratio is the main one

but even that doesn't mater if you have 5mi/5km.  The answer is not 1. L


You could do this on top of most languages

const num = new Number.Miles()

But that doesn't mean someone somewhere doesn't convert it back into a raw integer

And you don't get all the standard operators


but I was thinking if the compiler had an extendable library of unit conversions it could do the right thing.


Yep, doing it "in" the language itself at the compiler level would be better.


nods

and if you had an error you could define a conversion to fix the error.

then the assumption could be spelled out.


I like the idea of

const distance = 5 miles;

const time = 30 minutes;

const speed = distance / time;

// speed -> 10 miles per hour

time > distance

// Error! No comparison operator between minutes and distance.

That would be amazing


exactly!


Instead of doing * 1000 * 60 * 60 * 24 you could just configure what units you want as output.

And have automatic localization for countries and regions


So then you could add a line that defined the formula and the compiler could do that math.

and the output box could have the local defined unit and just convert, as you say.

math happening at the web brower output box L

define formula miles = kilometers / 1.62;

and the compiler could be smart enough to know that km = miles * 1.62


Would only numeric types have units?

Would strings / objects / arrays have units?


a string could be xml or tcl data or csv


Ahh I like it


or a language

then if you have english input and french localization it could translate using ai

it could be cool to convert tcl{ tags} to <xml>tags</xml>

I am saving this conversation to a text file [en]. L

Player Portal in Mega Mappers

Codex Engine: Player Portal, Auth, Rules-as-MCP & AI NPCs (Pure Node-Tree Edition) Defining game rulesets (D&D 5e, Pathfinder, Cyber...