How We Generate Art

The Core Problem

Without explicit guidance, AI image models produce wildly inconsistent styles. Here's what happens when you just ask for "cyberpunk character portrait":

Without References

Photorealistic

Anime

Oil painting

3D render

Same prompt, four different styles. Unusable for a game.

With Reference System

Same style, different characters. Game-ready.

The Reference Frame Budget

Gemini accepts up to 14 reference images per generation request. This isn't a design choice—it's an API constraint that forces strategic allocation.

Three-Layer Reference System

1

Gold Standard (Style)

age1_b_rooftop.png

#0f0f23

#1a1a2e

#00FFFF

#FF0080

#f59e0b

Always included. This specific image was selected through our style calibration process—a 120-image interview that determined exactly which visual characteristics define "CyberIdle style."

Without this anchor, Gemini drifts toward whatever style it "feels like" generating. With it, every image inherits the same DNA: NES-era pixel art aesthetic, limited color palette, clean linework.

"Use this reference image as a style guide. Match its pixel art style, color palette (magenta #FF0080, cyan #00FFFF, dark teal #1a1a2e), pixel density, and overall aesthetic exactly."

2

Scene-Type References (Composition)

Different scene types need different composition rules. A cityscape wide shot has completely different framing than a character portrait or an action sequence. We calibrated 12 scene categories through 120-image interviews.

Cityscape Wide establishing shots

Alley Street-level perspective

Terminal Interior tech spaces

Action Dynamic sequences

Portrait Mood Character emotion focus

Cyberspace Digital environments

Interior Indoor spaces

Rooftop Elevated views

Tech Detail Close-up technology

Dramatic High-contrast lighting

Rain Weather effects

Warmth Amber-lit scenes

Each reference was selected from 4 interview options based on composition quality, style consistency, and how well it guides generation for that scene category.

3

Character References (Identity)

Major characters get a 4-reference set. Here's GG's complete reference allocation:

GG's 4-Reference Set

PORTRAIT Primary reference

FACE Close-up details

BODY Full figure

ACTION Dynamic pose

Major Characters

4 refs

Portrait + Face + Body + Action

Supporting

3 refs

Portrait + Face + Body

Minor

2 refs

Portrait + Face

Background

1 ref

Primary portrait only

Real Generation Examples

Here's exactly what goes into generating different types of images—the input scene, all reference images, the complete prompt, and the output.

1

Age 1 Building: Street Hacker Setup

Building sprite • No character refs needed

Scene Description

"A cramped street-level hacker den with exposed wiring, multiple flickering monitors, and makeshift tech equipment. Age 1 starter building vibe."

References Used (2 of 14 slots)

Gold Standard

Terminal Type

Complete Prompt to Gemini

Use this reference image as a style guide. Match its pixel art style, color palette (magenta #FF0080, cyan #00FFFF, dark teal #1a1a2e), pixel density, and overall aesthetic exactly.

SCENE: A cramped street-level hacker den with exposed wiring, multiple flickering monitors, and makeshift tech equipment. Age 1 starter building vibe.

FRAMING: Square composition for building sprite. Clean edges, no character subjects. Focus on environmental detail.

STYLE CONSTRAINTS: NES-era pixel art. No photorealism. No gradients. Clean pixel edges. Dark background for transparency cropping.

Generated Output

Resolution: 2K Aspect: 1:1 Cost: ~$0.04

2

Story Hero: The Sick Day

Story banner • Atmospheric scene with silhouette

Scene Description

"The Sprawl during a brutal storm - lightning cracks across the skyline while GG's silhouette watches from a window. Cinematic ultra-wide composition."

References Used (4 of 14 slots)

Gold Standard

Rain Type

Cityscape Type

Dramatic Type

Generated Output

Resolution: 2K Aspect: 21:9 Cost: ~$0.04

Queue-Based Generation System

At scale, we don't generate images one at a time. Jobs are defined in JSON queues and processed in batches—either sequentially for safety or in parallel for speed.

{
  "jobs": [
    {
      "id": "age1_building_01",
      "prompt": "Cramped street-level hacker den...",
      "output": "docs-site/public/images/buildings/age1_hacker.png",
      "refs": ["gold_standard", "terminal_scene"],
      "aspect": "1:1",
      "resolution": "2K",
      "status": "pending"
    }
  ]
}

Art Interview Philosophy

Why do we use multi-round interviews to calibrate art style instead of just writing detailed prompts? Because visual art direction exists in a space that language can't fully capture.

The Problem with Text-Only Prompts

The Solution: Selection Over Description

Instead of trying to describe what you want, select from options. This is the same approach we use to prompt Gemini (reference images) and to prompt the human art director (interview choices).

Prompt Construction Pipeline

User scene descriptions get transformed through multiple enrichment stages before hitting the API.

STYLE_INSTRUCTION="Use this reference image as a style guide.
Match its pixel art style, color palette (magenta #FF0080,
cyan #00FFFF, dark teal #1a1a2e), pixel density, and
overall aesthetic exactly."

# Add framing based on aspect ratio
if aspect_ratio in ["21:9", "16:9"]:
    prompt += """
FRAMING: Wide cinematic composition with horizontal emphasis.
Key subjects should have breathing room. Use the full width."""

# Add character consistency instruction
if char_refs:
    prompt += f"""
CHARACTER CONSISTENCY: Match their distinctive visual
features EXACTLY as shown - same face structure, hair,
clothing, augmentations, and distinctive features."""

The Iteration Process

Major characters go through extensive calibration. Here's Cyber Chomp's reference development:

Round 1 Face Variants

A

B

C

D

Selected: Option C - best eye rendering

→

Round 2 Body Variants

A

B

C

D

Selected: Option B - proportions match face

→

Round 3 Action Variants

A

B

C

D

Selected: Option A - most dynamic pose

→

Final Reference Set

4 references locked for all future generations

This interview process was 52 iterations for Cyber Chomp alone. The investment pays off: every future image of this character will be consistent.

Strategies by Art Type

Type	Refs	Aspect	Strategy	Example
Character Portrait	0-1	1:1	NES pixel locked, contextual background
Story Scene	3-5	16:9	Multi-ref: style + chars + location
Hero Image	2-3	21:9	Ultra-wide, MUST center subjects
Building Sprite	1	varies	Transparent bg, isometric, NO rain
Resource Icon	1	1:1	64px readable, dark teal background

Color Palette & Style Rules

Technical Implementation

For practitioners building similar systems:

{
  "contents": [{
    "parts": [
      {"inlineData": {"mimeType": "image/png", "data": "...base64..."}},
      {"inlineData": {"mimeType": "image/png", "data": "...base64..."}},
      {"text": "Style instruction + scene prompt + framing + character notes"}
    ]
  }],
  "generationConfig": {
    "responseModalities": ["TEXT", "IMAGE"],
    "imageConfig": {
      "aspectRatio": "16:9",
      "imageSize": "2K"
    }
  }
}

Results Gallery

A sample of outputs demonstrating style consistency across different subjects and scenes.

All 596+ assets share the same visual DNA. Characters, buildings, resources, and scenes all feel like they belong in the same universe.