How does LukeW character maker work?

Since the late nineties, I've used a green avatar to represent the LukeW brand: big green head, green shirt, green trousers, and a distinct flat yet slightly rendered style. The Character Maker lets anyone create new versions of him, and the hard part is making sure every one actually looks like him. I tried LoRAs, ControlNet, IP-Adapter, character sheets. None of it worked well enough to consistently render assets the right way. Until now.

So what changed? It starts with a highly capable image model, but there's more to it than just picking an image model. When you type something like "playing guitar," the request goes through a few steps before any pixels get made.

First, a text-based AI model acts as a creative filter and enhancer. It rejects inappropriate content, interprets ambiguous words contextually (so "horse" becomes "riding a horse" and "coffee" becomes "drinking a cup of coffee"), and adds creative details like poses, props, and environmental context. Then that enhanced prompt gets wrapped in a detailed character specification with exact colors, outfit details, the signature three-fingered hands, and style directives like "full-body vector illustration."

But text alone can't guarantee visual consistency. The key insight was to load reference PNG images of LukeW and pass them to the image generation model alongside the prompt. Think of it like handing an illustrator a stack of model sheets instead of just describing the character over the phone.

Last, there's a verification step. Google's image generation API ignores reference images about 10-20% of the time. The validation step checks when that's happening and re-renders images when needed. Specifically, the image model targets a specific aspect ratio of roughly 1.75. If the output doesn't match, the system automatically retries up to three times. (And yes, I built and integrated the software using Augment Code.)

I've kept updating it since launch. People have made over 4,500 characters, and in March I updated the image model, error-checking, and prompts, moving to Nano Banana 2 because characters have more depth, better lighting, and more active poses. I updated it again in May with a new image model released that morning.

This connects to the design systems topic from earlier. The guidelines, reference images, and prompt rewriting together form what I call a steering layer: the steering layer is responsible for enforcing an AI image model's output aligns with the LukeW character's style, colors, appropriate use guidelines, and more. Swap the model underneath and the character still comes out right.

That's really the point of the whole project. Not just guidelines. Not just assets. But Tools.