GalCiv IV: Supernova Dev Journal #13 - AlienGPT

GalCiv IV: Supernova Dev Journal #13 - GPT

GPT stands for “Generative Pre-trained Transformer” and since the release of OpenAI’s ChatGPT the world of generative AI has really taken off. For the past several months, we have been working to build GPT technology into our Tachyon engine, Stardock’s cloud-based metagaming technology that has previously been seen in Ashes of the Singularity, Offworld Trading Company and Star Control: Origins. Thanks to OpenAI’s ChatGPT API and Tachyon, we are proud to demonstrate what this technology can do for games.

AI generated civilizations

In the past, Stardock would release new civilizations for $4.99 for 5 as a DLC. Creating a new civilization was a lot of work. Lots of writing and lots of new art and game balance. With Galactic Civilizations IV: Supernova, it’s just a matter of entering a description of your civilization and pressing a button:

Step 1: Enter a description for your civilization.

Step 2: Press the button.

There is no step 3.

It’s not just cosmetic. How you set up your civilization will affect how other civilizations will react to you and interact with you.

Let’s jump into the game:

I should also mention that when you save your civilization you can also play against them in the future… Such as…

Don’t judge me.

And other civilizations will refer to your civilization as you would…well, expect.

We are, after all, sentient ice cream.

And the Drengin are never impressed.

AI generated quests and challenges

Why stop with AI aliens? Another source of DLC in games has been new quests and related content. With Tachyon, our games, starting with Galactic Civilizations, can just ask for a new quest. Where is gets really crazy that we can ask in the language of the player so the result comes back as a native speaker (or alternatively, as good/bad at it as English).

Sharing

Players can take their custom civilizations and share them with others online. So we expect to see…well you know deep down exactly what we are likely to see.

What about the art?

As mentioned in a previous blog, Stardock has been making space games for 30 years. So we have a fairly extensive library of aliens to train AI image generators on.

The Human future

Ironically, this work has resulted in us putting out the call for even more artists, writers and editors. While on the surface, this may seem counterintuitive, let me walk you through how this works out.

Before: You hire artists, writers and editors and produce N assets per month which is insufficient to be commercially viable. I.e. the consumer market just won’t pay enough to justify focusing them on these tasks.

Now: You hire artists, writers and editors and product 100N assets per month. Now it’s enough to justify the work. The stuff the AI generates is really good and getting better all the time, only a human being knows our game well enough to know whether the output fits in with what we’re trying to do.

So the short answer is, we expect to hire more artists and writers and editors in the future.

We’re just starting

AI generation helps us greatly increase content but it won’t make a bad game good. It does, however, make the game a much more personal experience for the player.

We are hoping to make Galactic Civilizations IV: Supernova the most immersive space strategy game ever. Combining Stardock’s inhouse AI capabilities with generative AI we can build a truly living universe.

During the early access period for GalCiv IV: Supernova we will be balancing the game, adding more content, optimizing and polishing and listening to your feedback. It’s going to be quite a ride!

______________________________________________________


Galactic Civilizations IV: Supernova Dev Journals

95,700 views 31 replies
Reply #1 Top

Thats great news!

But whats about the citiziens portraits? generate to or use humans as citizien?

Reply #2 Top

 The problem is the pictures are static.Making it not consistent with the base races.

Reply #3 Top

Quoting wolfz1138, reply 1

Thats great news!

But whats about the citiziens portraits? generate to or use humans as citizien?

Since the citizens are actually code with pictures, that's a bit of a different thing.  We'll probably look at having the art for those generate eventually and have the behaviors (bonuses) be human since they're the safest.

 

Reply #4 Top

Quoting Ashbery76, reply 2

 The problem is the pictures are static.Making it not consistent with the base races.

We're actually fairly close to having these be animated too.  Just not ready for early access.

Reply #6 Top

Quoting pixelcowboy, reply 5

Is the early access released?

Yep.

Galactic Civilizations IV: Supernova on Steam (steampowered.com)

Reply #7 Top

Hey, could you have some way of providing the Early Access discount on Steam? I supported you Early Access but really prefer Steam much more as a platform. 

Reply #8 Top

It does feel like Early Access supporters are being 'punished' for what your business dealings were. Even though most people didn't buy and support you because the Epic Store. Zen pinball just transferred all my purchases from Epic to Steam after I supported them with Early Access on Epic, for example. 

Reply #9 Top

Sold! This alone is enough to sell supernova to me. 

Reply #10 Top

Quoting pixelcowboy, reply 7

Hey, could you have some way of providing the Early Access discount on Steam? I supported you Early Access but really prefer Steam much more as a platform. 

We have no way to do that I'm afraid.

Reply #11 Top

Well, I literally just went through this with Zen on PinballFx Early Access and it was certainly possible. You sent an email to their support with your epic account or proof of purchase and they sent you a Steam code. It is possible if your respect your Early access supporters. 

Reply #12 Top

Quoting pixelcowboy, reply 11

Well, I literally just went through this with Zen on PinballFx Early Access and it was certainly possible. You sent an email to their support with your epic account or proof of purchase and they sent you a Steam code. It is possible if your respect your Early access supporters. 

Ah so it was a manual process and not on Steam.

We can look at having some sort of discounted path for people who have the base game on Epic but we don't have anything right now set up for that.

Reply #13 Top

Please do. I want to pay the upgrade price today on Epic, but really don't want to own it there for all the platform benefits. I'll either buy your game in 2-3 years once it costs me $20 to buy, or today if there is a path to own on Steam for a comparable price. 

Reply #14 Top

Quoting pixelcowboy, reply 13

Please do. I want to pay the upgrade price today on Epic, but really don't want to own it there for all the platform benefits. I'll either buy your game in 2-3 years once it costs me $20 to buy, or today if there is a path to own on Steam for a comparable price. 

Thanks.  I sent this over to the IT team and they're brainstorming now.  I'll keep everyone posted.

Reply #15 Top

So I picked up the expansion based on this feature. (The other stuff added is cool too, but I really thought the idea of having AI generate the stuff for custom civs was awesome).

It seems to me there needs to be way, way, way, way, way, way more training for it. At least for the "image generation". I don't see how it's really "generating" anything when I see mostly the same images coming up for a civilization, regardless of what the description is. Super disappointed about this, but I hope it's just because it's early stages.

I havent even started a game because I couldn't get it to generate an image that was even remotely close to what I wanted. Even when I tried to make a civ of cats, none of the suggested ones had anything remotely like a cat (even though I've seen a few images in the past where it was basically a human with a cat head).

As for the text it generates, it's pretty good so far, although similarly, I think the AI needs to be trained on way more data because it tends to get repetitive very quickly (but I'm not really as concerned about the text description as much as the images).

Hopefully I can get over my OCD with the images and start a game lol.

Reply #16 Top

Quoting impinc, reply 15

So I picked up the expansion based on this feature. (The other stuff added is cool too, but I really thought the idea of having AI generate the stuff for custom civs was awesome).

It seems to me there needs to be way, way, way, way, way, way more training for it. At least for the "image generation". I don't see how it's really "generating" anything when I see mostly the same images coming up for a civilization, regardless of what the description is. Super disappointed about this, but I hope it's just because it's early stages.

I havent even started a game because I couldn't get it to generate an image that was even remotely close to what I wanted. Even when I tried to make a civ of cats, none of the suggested ones had anything remotely like a cat (even though I've seen a few images in the past where it was basically a human with a cat head).

As for the text it generates, it's pretty good so far, although similarly, I think the AI needs to be trained on way more data because it tends to get repetitive very quickly (but I'm not really as concerned about the text description as much as the images).

Hopefully I can get over my OCD with the images and start a game lol.

The image stuff is not the main point (since you can't really see yourself during the game). It will get better as more input goes in.  But it isn't the main point. That's not the GPT part.  

The GPT part is the text.  The descriptions of your civs and the conversations and quests and all the other content that comes from the AI.

I just tried "a race of cats"

 

This is what I got.  Did you make sure you were running as a carbon based civ?

 

TThe images don't get generated in real-time.  When you enter your description, the keywords get queued off to be generated for the future.  It picks ones that were previously generated.  So over time, more and more images will be made based on what you've put in. (for all I know, I got these cats because of what you entered in earlier and now it's smarter).

Reply #17 Top

Yeah, I tried again, and it finally had some that were cats. I guess you just have to keep trying and trying.

 

The text has been good so far. I just noticed some similarities because I kept re-generating, but again, I'm sure that will get better over time.

 

I hope more gets put into the images. Not just for the leader, but for citizens too. It would be awesome to have the race look how your leader is generated.

 

Also would be great to eventually have it generate ship designs as well, though I imagine that would require a lot more work than the text/images.

 

 

Reply #18 Top

So having picked up the early access and played with custom race creation a bit, some feedback.

I think you should have two fields one for the species description as in their character and civilization and another for their physical description.

Otherwise things you feed in physical description wise get reflected in the background story sometimes in an odd way.

The portrait generation is some what limited and often produces things nothing like your input, it does need training on a broader range of input, but obviously this is early access, but again would benefit if it's input was a separate physical description field.  As an example my race of genetically engineered human/wolf hybrids produced no pictures with wolflike features of any kind, I did get a tiger man.

You maybe need to source various animal images and train it with those as well as your game art as aliens might often be physically described by blending animals and humans.

It's a great idea though with a lot of legs and this is an early version, just thought I'd give some early feedback. I think most players will think the portrait important as well as the descriptions and traits.

It does like to make cat people and humans in robes an armour.

 

 

Reply #19 Top

Another example of a race I described as looking like a hybrid of humans and spides gave me a bunch of pictures of humans. It did well with my cultural description of them though.

Reply #20 Top

Oh I see so the images you see aren't generated by your input but from previous inputs, and it's searching for previous inputs tagged with similar key words. 

Reply #21 Top

So I'm only likely to get a portrait matching my description if someone else previously requested something very similar unless I keep trying until my description has reached the front of the queue? I guess I can see why you'd want to save on compute here especially as your player numbers scale up, personally I'd rather have the wait for output based on my prompt.

Reply #22 Top

Quoting econundrum1, reply 21

So I'm only likely to get a portrait matching my description if someone else previously requested something very similar unless I keep trying until my description has reached the front of the queue? I guess I can see why you'd want to save on compute here especially as your player numbers scale up, personally I'd rather have the wait for output based on my prompt.

you would definitely not want to wait. Think a 10 minute wait.

Reply #23 Top

I'd be happy having the option to wait, not everyone would be happy to wait, I would.

Ive been experimenting locally with Stable Diffusion already among other things trying to make portraits for GC4 so I'm well used to waiting and tinkering to get good results out of generative AI. 

Reply #24 Top

I also wouldn't mind waiting, if it was like 10 minutes. But then again, maybe it gets to the point where enough people seed the generated images where it wouldn't matter.

Reply #25 Top

Quoting impinc, reply 24

I also wouldn't mind waiting, if it was like 10 minutes. But then again, maybe it gets to the point where enough people seed the generated images where it wouldn't matter.

10min per image. So 90 minutes for all 9.

But at the way things are progressing, I wouldn't be surprised if we have real time animated generation before the end of the year.

6 weeks ago we didn't even think we'd have generated alien images. We were going to have that part just be the player selecting from the existing list of images.