When I worked on Manaburn, I worked with an extremely talented technical artist that figured a way to generate the assets for the game on his own. The team needed someone to work on the storyline, find a way to generate content via AI as well. I had the interest so I took it on.

I set one of our founders to work on prompt engineering and refining our prompting, with guidelines on how to make it more efficient. I worked on learning about AI, how best to do what we needed, at the lowest cost possible. I decided against rag training and fine tuning as we lacked sufficient data to train, and I did not think it was necessary to prove out what the team believed the core business was. I made a plan for advancement, and where our breakpoints were in the event we had to re-evaluate.

I systems architected the backend to provide the assets. Since most of the assets is to be churned through and burnt, there was no point in running up thousands of dollars worth of tokens to generate assets that would be thrown away - we decided to cache them all and re-generate based on traffic and # of users we needed to support.

I ended up porting the stable diffusion flow to serverless in order to demonstrate an end-to-end asset generation pipeline for 2023 GDC for the investors and pitches they had lined up. I took on creating a frontend meme generator in Typescript to showcase our backend for the Google Accelerator program in 2024.

The only portion I did not develop was the stable diffusion workflows. I wanted to tackle that to round uout my understanding of the complete tech for Manaburn.

I decided to experiment on Runpod, it was fairly non-commital and I wanted somewhere to just leverage GPUs to experiment with. I experimented with different prompts and loras to get the art I needed. As an engineer, we are always in desperate need of art to showcase our work and artists are exceedingly difficult to source (separate tangential topic on art and artists).

Out of consideration for Manaburn, I cannot post the process. Or, talk about the tech in any detail. But I can do a write up of results of my experiments and iterations.

The text is all AI generated, along with the art.

Banner art for my personal website - a mesh of San Francisco and Amsterdam canals. I would add my Singaporean pride in there as well but I doubt AI is able to at this moment. Generated off huggingface freebies. An image of me - pixar style. An image of me working - I am always working. Always need something green to spice up the UI - not professionally artistic, I will take the cheap wins.

I spent a good amount of time designing and writing up a few concepts for possible art use.