Automated image generation

[karlkablisk]

Thought we should have a thread about this topic.

I have been playing with this and wanted to note what to expect from what I've seen.

You can change how images are gereated in the options menu. I'll add photos when I get a chance so people have an idea of what to expect.

Wombo.ai option
This is the newest feature. Takes the load off your computer and generates pictures based off Wombo.ai its pretty fast but quality is a little off. I tried using the site directly and theres a big difference. But it's a great option if you need to relive your computer of the hard work. Even with a strong computer, of the game itself is demanded due to a large dataset, this can be useful.


From here are the options that use your gpu power

Quick
Pretty quick images are still surreal looking but it is quick. I did a default Atlantis game to test and it looked pretty good too. I think if your game has strong. Colors and asthetics like Atlantis this actually might be better than the other options.

Milddle option - balanced
Its what you'd expect, slightly better I dont have too much experience with this. Ill update this part when I do.

Quality
This one can look pretty photo realistic. With some ai shinanigans of a sudden cut in the picture.

-----
That's it for the options, now for custom.

There's a few boxes you can add numbers. The options above besides Wombo only automatically set these numbers for you. It provides a good guideline if you want to tweak it somewhere in between or beyond.

This is what I've gathered from experimenting and discussions so far on here.

Resolution
Height x width
Creates the Pixel density just like your computer screen.

Height and width as far as I can tell is a ratio. Keep to the ratio and you can scale up or down. I tried 2x the high setting. It was slow but worked. I didn't notice an improvement but it probably is better. I didn't test too much since it was too slow to want to keep on. I strayed from the ratio and it didn't load for me. Note sure if that's a hard rule. This does not need to be high to work if you look even the quality option is rather small. This is important to balance speed to quality.

Iterations
The ai will draw in passes. Like when you watch a printer print. If you say 10 the ai has to draw this in 10 "moves" kind of an equivalent to a human doing brush strokes.

Increase this and the ai will take several more passes at the image. As I understand it. The ai draws in "moves". For a human it would be something like brush strokes to finish a painting. I'm thinking when you see an image slowly develop on wombo.ai you are seeing iterations live. A high number means it can make fine details.

More iterations will increase image generation time. But also potentially increase quality.

Rate
I asked the dev in another thread where he gave more detail. I'm trying to explain it simply.
As I understand it something like this:

Difference between each move. A high rate low iteration will allow it to compete the drawing in less moves.
High iteration low rate will let it complete the drawing with fine details.

So think of these settings as a pair just like the resolution. Iteration @ x rate.


Using the options consider the balance of quality to speed.

I only recommend the higher image quality on a strong GPU. Otherwise you'll wait a long time for each picture and it'll bog down the game.

If you just want it fast and light weight pick wombo. If the pictures are starting to bother you experiment with the notes above.


Troubleshooting
Sometimes images will stop loading on any level.
From what I've learned from other threads, the log on the left side will show if Wombo is working. So you can see if it stopped there. And on the right is if it's gpu so you can look there.

Images run asyncrously to the other ai work. I think it's not supposed to take priority over text generation but I definitely see them working at the same time. The image Gen currently gets in the way and takes resources away from the ganeplay "turns" of text generation.

If you find you aren't generating images no matter how long you wait. Close the gsme and open again, that usually works. Maybe change image Gen methods before restarting to give it a jolt. That works for me.

That's what I've been able to gather about image Gen comprehensively.
I'll add screenshot of each when I get a chance.

[AI Roguelite: AI-generated RPG]

Thanks for the tips!

There are bugs with the image generation especially the wombo one just crashing and most of these bugs will be fixed in the next update.

One more thing to add: When I was experimenting, the image resolution influenced the quality of the image much more than I expected. For example, I thought a small image with 200 iterations would look more realistic than a large image with 100 iterations, but it turns out somehow the largeness of the image being worked on makes the AI just draw it better (in ways that far surpass the resolution itself)

For comparing with the wombo website, which preset worked best for you and what prompt did you test with? There's no difference in the way it gets the images so maybe I need to rework the prompts. The only other difference in quality would be the fact that it only generates long vertical images which will look stretched out in the game

[karlkablisk]

Above I did a breakdown of how I've seen it work in my experience so everyone can experiment with a good starting point.

Thanks for commenting, interesting notes, I'll experiment with that in mind.
To answer your question I'll give some examples of each level. Good for everyone to see.
By prompt I guess you mean the text I wrote, I'll note that below.

There are screenshots of the game in this community but mines specify the image conditions as the focus.

For all of these I went with a realistic example of a Feudal Japan setting. I started saving my world presets in an excel spreadsheet / google sheets so I can redo one if needed or if the word ends up not generating due to an error (usually when I push things too far by trying the 16gb model, doesn't work for me every time even with my computer).

And my 2nd example set will be more fantasy which will give different kinds of results.
My computer is RTX 3090, not an average example but I think I can give the full spectrum of low to high end expectation when it comes to the image generation in one report.
=============================
Text I used:

Starting point
My Castle Honno-ji. I am in the throne room in front of my vassals and wife.

World
It is Japan in the feudal warring age. June 21, 1582. This was the Sengoku era in Japan. There were many states at war with each other. Oda Nobunaga has the biggest territory and enemies on all sides.

Your desc.
I am Oda Nobunaga. I look like a Japanese Daimyo with a topknot drawn in a traditional Japanese art style. I am wearing Daimyo clothes of bright colors.

backstory
You are Oda Nobunaga, the Daimyo of Japan.
Your wife's name is Kicho.
You have a vassal named Akechi Mitsuhide.

=============================
There might be some unneeded details in each I don't have that down to a science (Yet).

Wombo
https://steamcommunity.com/sharedfiles/filedetails/?id=2841969723
In this case everything looked correct, recognizable samurai etc. When the concepts were more vague though, the art got real freaky and psycadelic.

It draws cartoons. Not pixelated more like water paintings. Strangely, this was the only one to give me a decent character portrait out of this Japan test.
Wombo feels very different from anything the computer GPU image generation will do.


It looks like a face in Japanese painting art style its a really good result for a random AI picture!

Load speed: This has the best load speed, it can really work concurrently with the AI doing its other work since its using the network download speed and a remote Wombo AI to generate the pictures. Its really nice if you want the images to go quick.

Quick
https://steamcommunity.com/sharedfiles/filedetails/?id=2841969559
Pixelated like a blurry picture but you can make out people, somewhat accurate.
My character portrait which I'll note separately from all the other images generated.
In general with GPU my character portrait pictures were so different from wombo I am wondering whats going on with the image generation or character portrait specifically. More than any other picture I would really like to "reroll" this since I'll be stuck with it all game.

For this one, I don't know whats going on some kind of scene of a woman running? haha.
Maybe if there was some context where it should know this is supposed to be a picture of 1 person. Unless people want it to be open to take picturs of anyone. Maybe theres some wording that can work in that staid. "this is a portrait of..." "I am sitting on a chair alone". Maybe some suggestions like that will help guide it and it won't need any kind of nudging.

Load speed was quick as the name suggests.

Balanced
https://steamcommunity.com/sharedfiles/filedetails/?id=2841969647
Pixelated but much better than quality. Like going from nes to snes. Still realistic photos of places and people but not all that detailed.
My character portrait was good. A man in feudal Japanese gear standing in the grass. Kind of semi action pose, really nice. Same description as the others, totally different result. I'm starting to think I should explain the scene and maybe even art style as much as the description for the portrait. I'm still not sure what words to exclude like "I have a" and character name etc. I usually put the name in here and the game catches it but not sure what it does the portrait.

Load speed: was a little longer but definitely fast enough to load everything as you are reading descriptions and getting your Barings.

Quality
https://steamcommunity.com/sharedfiles/filedetails/?id=2841974282
I tested this one more than once and generated dozens of pictures. So the notes below were consistent even with different game scenarios.

Photo realistic AI drawings. I say AI drawings because sometimes it will be random (like my character portrait lol) and sometimes it'll cut someone's head or be a zoom in of someone at a weird angle for no reason. But it draws really good things. An army of people, a real looking Japanese castle, Japan interiors etc. And if it doesn't look real it looks like a 3d model or a high quality painting like renaissance era painting but maybe even more detailed. (That might be a good keyword for the AI to if you want your pictures to be more realistic looking, "renaissance era style" worth testing another time.)

My character portrait... why did they do this? Its bad, real bad its either random or the AI really is thinking differently only here. I didn't take the screenshot with steam on my first try but it was funny and totally off the mark but it was a portrait picture of a face at least.

Load speed: Its slower but still workable and worth it every time. I think other computers it will be slower, this is the only level I think that will be heavily influenced by video card power so only use if you have the power for it. On my computer its still around the same speed of balanced believe normally it will be quite slow.

Now we are getting into custom territory
CUSTOM - 2x quality
Resolution 800x500 :: 100 iterations :: rate .2

https://steamcommunity.com/sharedfiles/filedetails/?id=2841982294
I only changed the resolution to 2x what I saw in quality so that the ratio would stay the same.

Don't try this! I suggest picking a level, quick, etc and then increasing iterations if you want to adjust things for some custom results. I'm not ready to play with rate, I think leaving it at .2 is the safe choice. Sometimes this doesn't even work. So I'll close and load the game.

I'd say it looks more detailed than quality, but not 2x like you'd expect. I'm seeing EFFECTS in pictures at this point that I didn't see in quality. So you'll see light effects more small details, take a look at "library" in the example. Theres 4 separate OTHER PICTURES in the picture. Fire lighting effects, light from a doorway. Really nice stuff. Loads more people in the army picture. Not sure why the location picture isn't as detailed,
Quality was definitely enough.

Portrait is... GOOD! Its just a close up of clothes but it looks good! Its correct japanese clothes, I included the description in the pic here.

I think the best way to use this is to increase the resolution higher on ONE thing you want more detail maybe the location, and keep the rest on quality.

Load speed: SUUPER slow,


Other notes

Here is the load game screen. This is good because it shows a few portraits and locations at different sizes. Note the locations since they are sized down to what you'd see on characters etc.
https://steamcommunity.com/sharedfiles/filedetails/?id=2841983140

I tested quality on a cyberpunk themed game it was same, super detailed with places that looked like real clubs, streets, alleyways, office buildings etc. People looked good it focused a lot on clothes and made real good looking clothes. Faces looked like real faces though they usually would block it with hair or something but nothin distorted at all.

I think I'll stick to quality, I like things looking actually looking like their descriptions with 98% accuracy.

Final thoughts
I think if we had a toggleable, image settings details that appeared at the bottom of the screen and people knew to turn it on before taking a screenshot, it would help identify quality patterns among all players without having to specifically do something like this thread. We could just compare screenshots and always know why things look the way they do. Good for the dev phase I would think.

[karlkablisk]

I have to make a note, wombo.ai images when you choose a image style is GREAT.
That nobunaga example above, I tried it with wombo's ukiyo-e style, all the art was amazing and super fitting.