Scrap Age: Survivors (SD1.5 workflow)

During my time at Visual Purple, I had the opportunity to work on this awesome Vampire Survivors-style game for mobile/PC. My main focus was collaborating with our Art Director Nic and fellow senior artist Shane on the icon work, where I got to push the boundaries of AI integration in our pipeline.

I chose ComfyUI as our primary tool since it was blazing fast compared to other local UIs for Stable Diffusion at the time. One of the cool things I did was create custom LoRAs using Kohya, training them on screenshots from our Scrap Age VR assets. Combined with some handpicked stylized LoRAs and checkpoints from CivitAI, we had a solid foundation for our paint-overs.

Our workflow evolved into something really efficient: Nic would start with rough concept scribbles, then I'd either build basic 3D elements, work them in Photoshop, or jump straight into txt2img generation. From there, I'd iterate through img2img until we hit something promising, then refine it with traditional paint-overs. Everything got organized in my PureRef doc, where Nic and I could bounce ideas back and forth with feedback and additional paint-overs.

I've been fascinated with image generation since the early days when it could barely make anything beyond dog noses. Being part of a small team at VP, I saw Stable Diffusion as this amazing opportunity to speed up our asset creation. After diving deep into the diffusion process and how these images are actually created, I came to appreciate it as another tool in our arsenal - one that raises the baseline quality for all art. Sure, there was some initial pushback both online and in the office (it was still early days for SD), but seeing the results speak for themselves made it worth pushing through those early awkward conversations.

As stated on the steam page:
https://store.steampowered.com/app/3143100/Scrap_Age_Survivors/
*AI was used in the BlueSky/Concepting phase of development to help ideate some designs of characters and UI to save time on our small team. Stable Diffusion using in-house created libraries and loras created from our own original art and not publicly available libraries was used in this process. These concepts are then taken and turned into functional useable game art created by hand, by real artists. Oftentimes this human created art ends up looking nothing like the original concept. No where in the game is any raw AI art used.*