Fireside Chat: Supercharge Your Game art Production - Volkan Gurel, & Andriy Matviychuk.

[Music] [Music] [Music] I would like to welcome to the stage to take part in this fireside chat our speakers to we're going to talk about supercharging your game art production and I believe we have uh beru from Leia Ai and Andre from triple dots is that right so please join us on stage coming up

yeah good morning everybody thanks for hanging in there uh this is Buju I'm the co-founder and C of layer AI we are a productivity Tool uh for game art teams our vision is to accelerate and unlock human creativity uh I'm emphasizing human here because layer is all about a tool uh to help art teams go faster to be more productive uh and we've been adopted by number of game top game companies already like zingga cplay triple dot of course is one of them marketing team of dream games um but yeah triple dot has been our very first customer uh that's why it's a very special moment for me uh like since the beginning of our journey like we've come so far and we really are excited to tell you all about this journey uh hoping to contribute to your own uh Exploration

with AI and Adoption of AI in your art pipelines to just go faster um well I'm accompanied by Andre Andre would you like to introduce yourself uh so I'm Andre I'm VP of creative at triple do been with the company for three years now and yeah last year was uh really interesting year when we started to work with ler yep well thanks for your uh early trust um let's just dive in uh we're going to keep it real we're not going to sugar coat anything world is not all pink so um let's kick this off um Andre so how did you first um yeah sure um how did you start exploring with AI were you looking at many tools like how did you decide to go with uh layer yeah so historically uh on our games for um um mainly for um your current content production um we used Outsourcing Studios

um and one of my directors uh a bit more than a year ago said uh is there any AI tool out there uh we heard a lot about it uh at the beginning that we can use for this content it's content that we create and throw away in a week uh is there anything that we can do better there so that really clicked uh with me and I started to look into some tools uh we looked in tools like um Mid Journey um stabil diffusion all that kind of stuff um and yeah at some point we realized there is something there um we managed to reduce um production of some of this content from uh two weeks to like a couple of hours now which was great mhm um yeah you touched on M Journey like we hear this very often what's the difference between mid journey and layer what was the problem uh there yeah so uh where while this was all exciting and um you know uh time saving and cost saving um we realized there are a couple of problem problems um first one was uh the

consistency of the assets produced uh it is great for uh creating concept art or get some ideas that we can work on but if you need to have uh production ready assets it was not great uh we had to uh over engineer the prompts and like try to have a library of proms that we can use for each parts of the different games um where layer is amazing in the sense that you we are able to train our own models based on our own art and you know training your own model sounds really really complicated um do you think it's really complicated to do it on layer no it wasn't actually it sounds complicated but it's not uh we the first test we run um on one of our games triple tile was literally to take um all the outsourced art that we had and trade the model on that basically uploaded uh 20 images um put some keywords in it and we already uh already saw the magic happening there how long does it take to train a model um around 40 50 minutes

and then you're ready to go every four or five seconds you get a couple of images generated there yeah fantastic so you mentioned triple tile one of the top games of uh triple dot what was the use case you found uh for layer first uh so first use case was uh so for um we have thousands of levels in triple tile and every 10 levels you have a new chapter so it meaning like a full screen background and six isometric assets um so that was the the first use case and we said let's jump straight into this and move into production change our Pipeline and see how it goes mhm well changing your pipeline wow like I can see gaming Founders here like how how am I going to convince my art team to change their pipeline so can you tell us a bit about like how did that change go uh so I think I should also mention before this um another problem that we had with M journey to similar to that is um uh the ethical part of it uh which

most of the art team was really concerned about uh me included as they should not going to lie um so uh being able to train our own model resolves that problem as well MH um and after that uh the our team really adopted this Tool uh across all the locations we have and um starting from triple tile we moved to other games M um I'm going to pause there for a second and talk about the copyright part so um every big AI model has a has a copyright problem because they need to be trained on this m billions of data points and that what makes them so powerful however there's no mechanism in place to reward uh artist for their contribution and that's really the main main problem and with tools like M Journey or stable diffusion based model there are still issues and we're continuously you know observing them but at later the way we use the these open source models is um

we allow game developers to train these custom models in a private workspace so during this um private like fine-tuning application uh no data is being passed to the Bas model provider none of your prompts or your assets are shared therefore you kind of get the world context power of stable diffusion but you com you make it uh really a handy tool to produce more art in your art style uh and I'm just going to go back to that so what was the benefit of layer to your art like artists so we talked about you know leveling up the playing field and you train these models of course in very high quality art by your senior artists so uh so yeah a couple of things happen on that um so one uh which I'm really proud of we managed to move uh our waro art team who is uh developing um liveops

art for our games and they were doing the same thing all over again uh for many years they wanted to progress they wanted to do uh UI other stuff but of course you know recurring content is that what keeps our uh games alive so we managed to unlocked the team completely we moved the whole livos operation using layer in uh Indonesia and uh all of those uh artists now are doing um what they love to do they do UI animation and they're able to work on um core game experiences instead of creating the same thing all over again mhm so you're saying that um now that models are trained with your top quality art by your senior artist Junior artists can use the same model exactly they can produce uh better quality things as well while uh learning to do that quality art as well which is phenomenal and how much faster did you end up going using layer uh yes so uh again if if we compare uh Outsourcing to when we used

mid Journey it was couple of hours now it's really a matter of a couple of minutes for the same amount of assets and how much cheaper is it in terms of time and resources uh so it's uh so the time we said it's it's we don't need to give an exact number but it's very cheap uh you know leadership is really happy about that Fant um we basically been able to close all out sourcing uh operations um up apart from one game that we're still figuring out um yes so um triple dot is using layer across multiple games and across multiple teams from you know game art to monetization Art to marketing art um and we also you know work across multiple games uh voku is one of them um I would say like backgrounds and in-game uh props are the top the top top use Cas yeah exactly exactly and uh with Piper Pet Cafe so like what's not working

there what do you think like we need do we need to refine uh the models a bit like yeah yeah yeah so um we have a particular perspective in Piper spit Cafe particular is isometric perspective um what we seen with layer sometimes we cannot get exact perspective we need um for some assets it works like um I know plants or trees but when it comes to chairs or um I know bed and things like that uh sometimes it doesn't get the perspective right but we might some progress there were a couple of versions of the of the model and we're still working on this I'm pretty confident we can get it so to really talk about that problem it's all about training the model uh with the right type of assets so if you have a really consistent angle we need to make sure that the model consist that angle because after all layer today is only supporting 2D but we're really stretching the limits of a 2d tool to act like 3D uh that's why there's some you know gaps in the consistency sometimes but You by working

with the team and you know we are all uh well trained to support customers we've been in games for a while I worked at Unity for six years doing customer services so uh with some refinement uh we're confident Piper Pet Cafe also will get there um so let's also talk about maybe um like was your team okay to adopt AI uh in the beginning in the beginning not at uh there were lots of fears on um you know ethical part also job security uh there those all those questions were raised um but you know we had a couple of sessions after we adopted lari uh all the ethical part uh disappeared uh there there were still concerns about job security but the our vision is to produce more with what we have already so we have same teams we're still growing still hiring but we can do just more and faster which is fantastic so um one of our goals goals is you know to

make sure that you can do more with what you have and i' would say in top gaming teams like like how big is your art team again for everyone's reference uh so we around UH 60 people in my team this is excluding Indonesia uh operation yeah and with no reduction in the team they're just able to ship uh levels faster you know games are scalable players are demanding but in the past they had this huge art bottle neck yeah yeah yeah yeah remember we uh we decided we need to have thousands more levels um we were just figing out okay we need more Outsourcing like we need more resources and everything but this all changed with layer yeah fantastic to hear um yeah I'm I mean maybe like we can even ask questions early on yes copyright questions I'm sure is going to come but yeah please go ahead yeah so uh I would like to know how you training the on your own make e on theone of

which trained onli of parameters from the internet so what do you think about that yeah great question and um unfortunately it's it's a tough question and it's a evolving space it is true that uh stable diffusions based model um has this problem um and we're building layer in a model di agnostic way if there's a better model that is let's say reward artists we're going to make the switch in no time currently the way we handle this is at least our tool is built on custom model training we're not so much promoting um stable diffusion based model usage in the end product um the way we use um layer is not passing data um it's not trained ethically but at least we're not contributing to his

further success we take what's great about it which is the world context uh in terms of like uh the text comprehension capabilities uh and then um using developers custom artwork like they produce more in their art style not so much purposefully trying to infringe copyright um we have some guidelines we work with top uh Law Firm from um Israel actually arog La um so yeah we're we feel we're responsible to at least educate uh the users about like what is this we try to provide guidelines we tell them how you cannot infringe copyright what words you should not use like how you can be respectful of people's personality rights so for example if you type Madonna you're infringing Madonna's copyright because table diffusion obviously knows who Madonna is or if you say Picasso or vangog because stable division knows

like you can easily uh make a wang go uh painting with stable diffusion and that's not okay so um the good news is though and I'm not going to give so much spoiler we're we think it's a huge problem and we think there's a huge need to solve this problem and we are working on something which we hope to announce uh hopefully addressing that problem yes please yeah the yeah so so the cool thing in lirer that we can if we have some artifacts uh there there's canvas that we can fix uh different parts of it uh right there without even needing uh to to put the assets inside Photoshop and edit it

there so so far for for our type of games and the type of assets um you know the quality level is is still high and we can fix things that are you know they're wrong um it's really nice we we have multiple layers in there uh we can select different parts of it we can prompt it again we can paint over and then render that again within the with the tool um so it's really easy fast yeah great point on that so while layer is you know powered by AI uh we try to keep most of our features similar to photoshop so uh we want to make sure that this is a tool for artists and AI is easy to adopt and familiar to adopt because there's so much new uh details about Ai and it might look confusing or complicated especially for it like a more traditional artist and maybe a technical artist is better off like faster in adopting uh these tools but you want to make sure that to bridge

this gap between an AI native person versus a traditional artist so um keeping the tool very familiar uh to photoshop has been very very rewarding you can adjust your brght softness you can build in layers you can switch between different AI models build a composition like really quickly you can export to photoshop uh also most of our users are exporting in PD format take taking the asset to photoshop to finish it h why because a ler is a startup it's only been a year how can we like be as good as Photoshop in edits that's why we made it super easy for them to take it to photoshop and actually we have many users in the in the audience like they're always saying like with this Photoshop exports it's been even more faster because at least we're not forcing artists to use an AI tool like constantly like we we realistic about it the point is can they make art faster is it fun to

make um can we make AI um an exciting tool that uplifts their productivity that's all we care about yep y please sorry use the back I was just wondering what size AET catalog would you need for training layer AI to produce good results yeah absolutely so um AI models are constantly evolving um the first model we start we enabled this custom Training on uh was trained on 512 by 512 uh pixels uh therefore it was defaulting those assets and you can use upscaler tool to go up to 4K because you know you may have like defaults um that you work at like higher resolutions uh but then we're uh enabling the newer like latest um model that was trained on24 by 1024 so you can have really high

quality defaults at this uh resolution as well and use the upscaler to upscale what's your Default Resolution uh different teams do uh different things uh but uh yeah at the end of the day we always upscale everything and maybe adjust while it's upscaled it gives more detailed result as well exactly you can make it more HD how many assets how many assets um to do you need to train a model so per game uh we recommend uh train few model mods for different use cases like backgrounds game assets like sorry in game let's say it's a decoration merge game like Furnitures backgrounds um characters it can be anything uh for each model it's 25 assets yeah in triple tole we have four models basically for for for the game um we have for buildings for trees for other assets and the backgrounds basically mhm yes go ahead yeah last question apparently but we can

always chat uh later yeah you have oh do you have guidelines about what kind of prompts artists use uh in general to use layer or each gaming company creates their own prompts and prompt Library uh for uh artists to be able to create characters or scenery um in an easier way cuz um even using chpt at first people didn't know what write to be able to get the right result so so are you working on um guidelines on prompts uh overall in layer or per game yeah sure so um great question and this is again one of the new skills you need to develop like how to command AI to get what you want um so yeah prompts are the text words you use to tell AI so like let's say a table with a way on top with the rows inside right um we have some guidelines but nobody reads them so we need to make confirm yeah exactly and um

most of our like our users are from everywhere um and they're not native speakers and the huge problem is like how how do we make it so easy we're working on certain uh improvements in the product that helps read their mind not read their mind but uh we have an enhancer for example you write a word it gives you a big paragraph so you can use that inspiration and edited uh but it doesn't need to be that complicated right um model also will get better at comprehending text uh and hopefully like get better at like giving you what you want with as few as word as possible but there are like several improvements where are we're working like suggestions uh prompt history um saved prompts um but these are all private to company's uh workspace so generally companies they work with you know multiple games multiple Studios they share insights within themselves we're not a consumer tool or you know you can go check out prompts if our users demand uh more

consumer stuff we'll do it but currently like everybody loves that this is a private clean workspace for their team to to share I think you might have to leave it there and Andre thank you very much thank you so much yeah thanks so [Applause] much