Public models, proprietary orchestration.
Our visual pipeline orchestrates a curated set of public foundation models — image, text-to-image, and video — in a proprietary sequence tuned for slot-symbol readability, brand consistency, and premium casino production values. The models are off-the-shelf. The orchestration, the prompting discipline, the visualStyle authoring stack, and the QA loop are ours.
Slot art has a specific set of constraints that general-purpose image models don't respect out of the box. Symbols have to read at 128×128 pixels on a phone. Brand palette has to stay consistent across 40+ generated assets per game. Character work has to survive resizing, layering, and animation. And the whole set has to look like it came out of the same studio, not eight different sessions on eight different models.
The orchestration layer is where the proprietary work is. We treat the public models as a set of composable primitives — one for concept exploration, one for symbol production, one for character work, one for motion. The sequence in which they're called, the structured prompts that drive them, and the human QA loop that catches drift — that's ours.
{
"visualStyle": "warm gold, sun-bleached wood, mining-camp palette",
"mechanics": ["hold-and-win", "expanding-pile", "codger-hands-you-gold"],
"mood": "playful nostalgia, not literal gold rush",
"readability": "symbols must survive 96px",
"brandTokens": ["#d4af37", "#0a0908", "#e9e5da"],
"characterRefs": ["codger-v3", "durian-mark-v2"]
}Every game gets a visualStyle brief like this before any asset is generated. It's read by both the human designer and the orchestrator. It's what stops game #4 from looking like a different studio's game #4.
The human QA loop is not optional
[Design.Human] runs a bug-and-task burn-down across every game, with a yellow-highlight system that flags anything new since the last review. It's how a designer catches a symbol that's technically correct but reads wrong at scale.
10 open tasks and 5 open bugs on HYG. Highlighting anything new since Tuesday in yellow so we don't lose track. Aiming for zero before we hand this off to [Sound.Human] for full audio.
The burn-down isn't a project management gesture. It's the mechanism that turns a big pile of AI-generated candidates into a shipped, coherent, playable game.
AI still can't come up with better creatives with writing and branding than humans, I'm afraid. Good.— [Design.Human], on where the AI ends and the designer begins
What we don't publish
We don't publish our model stack. That's not defensiveness — it's the same reason a restaurant doesn't publish its supplier list. What separates us from studios that wrap a public model in a wrapper is the orchestration, the discipline, and the human standing in front of the output saying yes or no. Those are the assets. Those are what compound.