Session: Studio Speaks: The Reality of AI Adoption in Games - Ilya Kuznetsov & Tomi Huttula.

[Music] We're going to start off by talking about the reality of AI adoption and we've got uh Ilia and Tommy coming up. So Ilia from Alia AI and Tommy from Cosmic Lounge going to come on up and tell us all about uh the reality of adoption in the studio. Thank you.
All right. Um thank you everyone for joining the session. Uh we're happy to kickstart the practical AI uh track and my name is Ilia. I'm director of BD at Layer and Layer is the end-to-end platform that empowers creators and artists uh in video games industries by

letting them to tap into AI models and scale their production pipelines and basically makes their lives easier. And as mentioned in the previous session, there's a lot of hype and buzz around AI. And it's very easy to get lost. And we know that it's true because even this year PGC has not just one but two tracks dedicated to AI. And we're happy to talk about the practical application uh and the real world examples how AI helps studios today's today. And to help with that we have amazing Tommy from Cosmic Lounge. Cosmic Lounge is an AI first puzzle stu puzzle game studio that has amazing examples how AI can u bring innovation and empower teams. I will let Tommy introduce say a couple words.
Yes. Hello everybody. It's great to be here um to talk about my favorite topic which is using AI for game development and um yeah so that was quick introduction about myself but maybe just to add to that that I've been in the

games industry for a long time and actually started we just we're talking about uh gambling and casino games that's where I originally came from but we have been doing mobile games for for ages as you can see from the long gray beard already. Uh so the company Cosmic Lance we have founded 2022 uh in March and if you think back about you know time um there was no AI boom then so it wasn't like a big thing um and and how we why we started the company it was founded by uh six game industry veterans who have been working on puzzle games for a long time and and we started from the angle of like what are the big problems like because we've been working for different companies and seen uh uh companies repeat same sort of mistakes or making their life uh miserable sometimes building games. So we wanted to solve those big problems we've seen in the in the industry and that's where the starting point was and and some of those big problems was like

how do you know if an idea is a good idea? So in game companies there's a lot of creative people. Everybody has an idea of a potential new hit game, but you don't really know if that idea is a good idea before you have some prototype to try playing that. And that that's uh we wanted to um make it as easy as possible for our designers to try different puzzle game mechanics. Then there's other areas that that are always um challenges for especially for casual puzzle games. the content uh some of the games they have more than I don't know 20,000 levels and and this content production has always been a limitation and that has of course limited designers to create games that they wanted to do. So putting all these problems together we started uh thinking about what are the solutions and we came to conclusion that if we can do our own engine which is uh making it uh uh giving no code tool set for our designers so they can try different things easier and then uh for content production and architecture

we had very different uh different sort of uh technologies that companies generally are using and that's when we started thinking about the AI we they were different AI models started popping up and we realized that hang on a second so if we put AI model for this use case for that use case and that use case we get a huge boost for the uh advantages that was already uh provided by our technology and that's I would say how we became AI native
amazing and you mentioned uh building prototypes there's the engine behind it right how is it called
uh we call it puzzle engine so we are not licensing it so it's it's just an internal tool but it it was uh it yeah one of the differences in that engine is that um all the puzzle game logic is in data format. So the designers actually use tools to configure how different components work. So that makes it very easy for them to come up with new ideas. So this is not not just about matchy but

any kind of u puzzle game uh mostly grid based but there's a lot of like unlimited u amount of uh great game ideas. um uh which nobody has yet done and that's what uh our tool enables our designers to uh test and try.
So there's one aspect of it would be testing game mechanics and trying some prototypes and letting your developers and the engine test the mechanics before you push it into the game title. Uh the other part of it is absolutely art, right? And so how do you currently navigate or how do you how do you use the technology nowadays to speed up the production test out new uh assets in the game and place them in the game? Is there any Well, I'm kind of leading here today to layer. Yeah. Can I I'll answer to that soon but I'll actually uh say something about the the testing of the idea. So we always test with real players. So the where AI and our engine is helping our designers to uh you know

sign significantly accelerate the production of prototypes which are like really good-looking uh production level prototypes that you can that we then test with with different methods with different players. So, and and actually related to some things that were mentioned in the previous um uh previous session was that AI doesn't really know what good looks like or what something new and good looks like.
Doesn't have a taste.
Yeah, it doesn't have a taste. Like I really like that that comment. So, so uh for us uh it is often about making things quicker so that we can test it real players. Then there's some use cases where you can tell AI what good looks like and that's one of them one of those examples is like level level design or level generation because you have sort of much sort of a narrower scope uh of you know what the good looks like and there you can there you can let AI make the make the recommendations that I think this is good but then going

back to the to the art side obviously that's probably the most familiar use of AI for for a lot of people um and and we started using um actually also when we started testing different art styles and different concepts we started using very early uh AI generated art for that just to sort of find out what what people would like and um so we've been using it from the day one I would say uh u and here I would I would also mention uh that we have we have an art team and we have very professional artists who have been doing game for long games for a long time and that applies for uh I would say for every every area of game development you still want to have an expert for each area but you want to give the best possible tools for those experts and and that's where the AI comes in and that's also applies for u for art uh so so the different use cases for art uh like there's the if we go back to prototyping that's where I mean

it it's like huge acceleration of trying different things seen visually even internally when you communicate things. Uh but then when you test with test with real audience absolutely a artisi is is fantastic thing. Uh then for the prototyping we wanted to make the workflows as efficient and and seamless for our designers. So when they do prototypes uh we actually integrate it to our puzzle engine. So same tool where they design the mechanics and levels we integrate it to different art models that they can from that UI actually generate the placeholder AI when they do the prototype and then later we actually uh now have also done a layer integration. I'll we can probably go to uh details why why we really like using layer as part of our production process. But uh for some 2D art uh I think in our current game which is in soft claws I think there's there are images which are

straight from uh layer so
directly into the end.
Yeah. So I think there's so it really varies like how much manual work do you need for certain art. It's up to the artist to decide but I think uh we are getting to a production level uh uh art in some cases already. Uh also um yeah maybe maybe
yeah before before we move to the art part and the integration uh I I I would like to emphasize what you mentioned before that replacing experts in the team is definitely not the avenue people should be and studios should be pursuing because uh AI is not out there to create IP or create original art. It's still up to the artists, their vision who define the art style and the vision behind the the game. And then what AI help can help do is replicate that vision, replicate that art style and then help the teams produce the assets at much bigger scale

than they would do without the help of AI. And I think this is where the so-called custom styles or loras are coming into power and play. uh when we can train AI to understand the art style to reproduce that art style in more inconsistent way so that yeah you can basically save a lot of time and resources while keeping your original vision in the game and I think we have one quick example to show here with uh
so these are uh two uh gifs or gifs I still don't know what's the right way to pronounce it uh that are showing the integration of layer into uh cosmic lounge's puzzle engine And maybe Tommy you can share a little bit more. I don't know whether we start with the bottom one or the top one. Uh where you can see a generation of assets and then going directly into the puzzle engine uh and the preview of the scene in Unity. Is that correct?
Yeah, let's let's do that. Um so yeah, the puzzle engine is built on top of Unity. So if you see some familiar sort

of UI components, that's that really is on Unity. There's a lot of code in puzzle engine. So it's actually quite huge uh piece of software. So what you see here is uh let's if we start from the top um so this is um showing the sort of prototyping prototyping phase. So uh here the designer is thinking about new kind of mini game and and uh generating uh ladybugs and whatever they butterflies and um the only thing what the what the designer needs to do is actually just write the ladybug because the uh the designer has already selected pre-selected the style which has been trained in layer. Uh, of course you can you don't need to limit per style, but you can and that has that is super important especially when you go into production. So here is um on the left you can see an image of the you know a short clip of our game pets and puzzles. So there uh there yeah what what is shown here is like you can actually

generate using certain style which has been trained based on images created by own artists. It comes straight from layer and you can pretty much pick and just drag and drop or just select a new asset and it and you can see immediately what it looks like in the game. So that makes it also very fast to sort of try out and see see what works. And uh if we so we covered the uh creating art scaling production using custom styles to create more consistent art and place it directly into the game using the API. Um there's one thing uh very there's one thing that the developers or artists sometimes struggle and that is creating the art uh assets in the required angle.
Yes. Yeah. Yeah.
And that's something that developers often struggle because AI doesn't really understand what is isometric view. It can skew it a little bit by a couple of degrees. So is there a way to go around that or not?
Yeah. So this is one specific uh problem like if you look at the game image it it

is a certain perspective and when you generate 2D images you never really know what the perspective you get. So it's a bit of uh trial and error. Uh so sometimes uh what we have started also doing is that if if you get some image that you really like but it's in a wrong perspective uh what we have done is using different uh arti tools of uh taking the like in this case uh maybe one of these wall bush the perspective doesn't matter but let's take the dog house so you take the 2D image which is art generated based on the style uh trained by our artists you take that into a a 3D tool generate the 3D uh put it in the right angle, take it back to a 2D uh uh AI uh tool and then you have it then you have the angle and you have the still the same style. So of course a lot of these sort of there's there's actually well um it might sound like that you know artist just presses a button and everything happens. There's actually a lot of back and forth between

the artist and different AI tools. uh but but there's a huge production uh sort of efficiency increase uh when you when you uh when you have the best tools and in in today it really means using tools with
actually that workflow uh would be possible in layer as well since layer doesn't only power 2D generation it offers 3D and videos as well so as you mentioned uh you create the asset you move it you use the image to model flow to create an asset in 3D model uh generator Then you place it in the right angle. Then the trick here is that when you regenerate that asset into 2D back into 2D, you're using the trained style which will bring back the asset visual art style uh into this consistent art style that we're trying to pursue. And this way you will not sacrifice in the uh in the direction or the art style vision. Uh it using this trick helps us to reposition the asset into the right angle which is a great life hack. I actually didn't know about this. Thank

you Tommy for sharing this. Um but now we have we're getting short on time. I would like to cover um the topics that are very popular these days and that is advertising in UA and uh from my experience what I see how studios are utilizing AI that is very much into video uh creating narrative based um videos for advertising or even testing some mechanics or mini games that haven't been out there. So uh Tommy is what's the pos or stance uh of cosmic lounge when it comes to you know creating UA videos what value does AI bring uh and what problems that helps to solve uh when we try to you know market the games ethically.
Yeah. So uh I think well all of us have seen the deep fake videos and like so you know that AI can generate you know it's getting pretty good in creating

videos which you can make look like anything but maybe starting from the mini games. So I think everybody's um familiar with misleading ads and and mobile gaming companies are using them a lot because you have a lot of freedom of making whatever kind of ad that attracts players and you get a low CPI. Uh but then the uh problem is that if the players don't get what was promised a lot of them churn but still uh enough players stay that you you have a lower cost per retain user. Uh so what we what we did with our engine was that um our uh designers are able to also create mini games using the engine. So um uh because of that we are not using misleading ads. So the as we are showing if there's a mini game it also available in the game and that's why we can avoid the the the um churn problem there.
There's an example we can show as well.
Yeah. Well actually let's maybe go back. Yeah. Well that that's fine. Let's let's stay here. So okay
I need to tell the background of of

this. Uh so um all right. So that's the mini games that we use and of course you can just record the video uh the mini game video and then you can add it to the ad creative and all that. But now now related to ad creative. So um why I wanted to show this image is actually from one of our company parties and every we have like two parties summer party winter party every year and we always do some fun fun experiment with AI and this is from the last one where the idea was that everybody had to come up with the worst possible miniame idea and and so then we were voting like who has the worst idea and the idea was uh then visualized with the with the video which was AI generated that's what you see on the left. And then to make it more fun, we just um u created a fun characters and actually recorded them like made them as AI personas and made them review those bad videos and that's where this this screenshot is from. I probably will put the whole video of all these characters to you know to LinkedIn

because this they are super funny. Anyway, so what what we found out there was that AI is so well trained making different characters and like how they act and how they tell the story. It's it's really good like you really feel like a movie producer when you work with some of the latest um AIS for videos and of course for ad creatives that's super super useful. But then the miniame video what we also found out is that when you were trying to um even if you uh we showed the sort of starting starting frame of the miniame idea uh and the end frame the the AI really didn't understand the game gameplay logic. We were trying to explain like this is what happens in the mini game. Make a video of that. But like it it's it's that was like one finding that AI doesn't really understand how the how games work yet. And that's something what of course we have designers for uh but our uh plan how to take this further is to uh when you have a new mini game idea in puzzle

engine then you can just pretty much ask AI to generate at create this based on uh based on those games. So there's lot of lot of steps if you combine the right technologies and and and um and tools and the workflows I think you can there will be a lot of better ways of doing marketing content uh in the future
and that I'm afraid is a fantastic place for us to end it. Uh thank you guys that was really really good. Uh we actually don't have enough time for questions unfortunately. So, but round of applause, please for the fantastic introduction. And I to I'm totally with you that, you know, AI doesn't understand games. Let's face it, AI doesn't actually understand [Music]