EverGen

The problem

Evergrove is a live game on a content treadmill: every update needs a new batch of buildings, ingredients, characters and backgrounds, and the volume didn't fit inside the time available to make it. The generation tools that existed weren't solving it either — ComfyUI and its family are powerful, but each demanded heavy setup, and none held visual consistency across a batch reliably enough to ship from. Every artist ended up with a slightly different setup producing slightly different results, and reconciling them ate the time the tools were supposed to save.

So the real requirement wasn't "add AI generation." It was: move at the speed the content calendar demands, and come out the other side with assets that look like they belong to the same game.

The system

I designed and built EverGen, an internal tool that took the whole path from prompt to game-ready asset and put it in one place, running on the Gemini API, shaped around the specific job the art team had:

• Generate: production buildings, ingredients, characters and backgrounds, driven by reference assets and a style prompt so output stayed on-model instead of drifting per artist.
• Curate: review the batch and keep what works, in the same interface that produced it, instead of exporting to a folder and sorting by hand.
• Clean up: background removal built into the flow, so what comes out is usable rather than a starting point that still needs cutting out.
• Organize: assets grouped by the content update they belong to, so a release's art is a set you can find rather than files scattered across drives.

The decision that mattered most was keeping all four steps inside one tool: generation was never the bottleneck, everything between a good generation and a shippable asset was (choosing, cleaning, naming, filing). Automating generation while leaving that tail manual just moves the problem downstream, which is what the off-the-shelf tools were doing. It also meant near-zero setup — an artist opened it and worked, rather than configuring a pipeline they'd route around once the deadline got close.

What broke

• Style drift: later generations in a batch quietly wandering from the approved reference, worst on long content runs.
• Off-brand output: technically clean images that were simply wrong for Evergrove's look (wrong mood, wrong palette, wrong level of detail).
• Messy edges on background removal, which turn into visible artifacts once an asset is composited in-engine at game scale.

The fixes were mostly about tightening what the tool held constant: locking reference assets harder, standardizing the style prompt so it stopped being per-artist folklore, and treating curation as a real step in the pipeline rather than something that happens after the tool is done.

The result

EverGen cut the art production timeline for Evergrove content updates roughly in half, and was adopted by the art team as the way that work got done — not mandated, adopted, because it was faster and more consistent than the alternative. A tool that produces good images in a demo is easy. One a team keeps choosing under deadline pressure, update after update, has to be right about the whole job, not just the interesting part of it.