Any game developer can tell you how challenging it is to create, iterate, and optimize creative assets. Games require a constant pipeline of new ideas to keep players engaged and compete in-market. This was the challenge Bunch was trying to solve in their digital world, giving players assets that let them be themselves - without compromise.
Who is Bunch?
Bunch is a mobile digital world where people meet, talk, and share anything, from watching videos to playing games together. Built by a team of 35 and funded by many large game companies, from Supercell to Electronic Arts, Bunch aims to create a virtual playground where players hang out before and after playing their favorite games.
What’s the problem?
Like many other games, creating a constant pipeline of virtual goods and spaces is critical for Bunch to keep their players engaged and build a vibrant in-game economy. One of the ways Bunch offers players a way to self-express is by allowing them to customize their avatars and virtual rooms and homes. This creates a constant need for new assets, which they can never fully address due to limited time and resources.
How does Gen-AI help with never-ending needs for assets?
With the developments in Generative AI, the Bunch team started looking for a way to expand their bandwidth for virtual goods and live operations- and turned to usual places for creating assets. While they were impressive at making art from scratch and driving inspiration, they still needed a tool which allowed them to collaborate and produce art consistently with their unique art style to speed up their art production.
What made Bunch choose Layer?
Bunch heard about Layer through their games network and jumped right on it, hoping to find a great tool to achieve the style consistency, speed, and simplicity they are looking for. After training their first styles for Bunchies and Bunch Homes, they were reached out by Layer’s Artist Operations Team, who suggested improvements to achieve perfection. After implementing the quick fixes recommended by the team, which we’ll share more about below, they were convinced that this is the tool they would like to adopt in their art pipeline.
Here are the steps Bunch followed during their exploration:
Deciding on what to make with Layer: Bunchies Avatars and Virtual Homes
Training style(s) by uploading assets and writing descriptive captions for each asset group.
Upload up to 25 Avatar Images and Homes respectively and Write Captions (describing what the image is about): Detailed descriptions outlining: material, lightning, gender, emotional state, outfits, poses, colors, hair, and distinct features. Choose the one that’s the closest to what you’re looking for.
Here we have blurred the inputs and captions. The important thing to notice is how consistent the images and detailed the captions are!
Decide on test prompts.
Example prompts are used the create outputs of style variations so that you can compare and select the best.
Pick a variation.
Select the best version of your style to use in future forges.
Test your style(s) by forging: Start prompting to create your desired assets and see the magic.
See your amazing new style in action.
What were the learnings?
Bunch took a couple of attempts to get it to the level of consistency they were looking for; they quickly realized adding the right type of images and writing detailed captions was key in getting the desired consistency. Here are some of the learnings:
1. Train different styles for different use cases: Homes and Avatars were part of the same training, which caused mixed results.
2. Vary your assets (avoid repeating images): Bunch at first added too many repeating images, which limited the creativity of Layer in early results.
3. Write descriptive captions: at first, the images were uploaded without descriptive text, which affected Layer’s ability to learn the style from the text.
4. Enhance your prompts: Together, the Layer Team and Bunch had the idea to leverage ChatGPT to generate detailed prompts. Bunch uploaded Layer’s prompting guide into GPT 3.5 and asked to expand the prompts, which led to even more impressive results. As a result of this hack, Layer integrated GPT 3.5 to enhance prompts when forging.
5. Upgrade to Enterprise Plan: to work closely with the Layer Team to achieve faster results you’re looking for.
Together with the Layer Team, Bunch was able to generate concepts that matched their style very quickly, allowing them to offer their players a great variety of avatars to pick from and a selection of virtual homes for them to hang out and play games together.
If you are new to Layer, come check us out; if you have tried before and haven’t found what you are looking for, we recommend coming back and trying again based on these insights!
Your presence fuels our innovation, and we can't wait to present you with more feature updates and case studies soon!