AI Character Reference Generation

AI Character Reference Generation

Create 17+ consistent character poses from detailed prompts — no reference images needed.

⚠️ Staleness Watch (Last validated: July 2025 — Flux 1.1 Pro; Feb 2026 — nano-banana-pro): Image generation models on Replicate release new versions and deprecate old ones. Before generating, verify the model identifier at replicate.com/black-forest-labs. Upgrade path: black-forest-labs/flux-1.1-pro-ultra provides higher resolution (up to 4MP) with the same API surface. black-forest-labs/flux-2-pro adds multi-reference editing. google/nano-banana-2 is the faster/cheaper successor to nano-banana-pro (same image_input API). Input parameter schema may change between model versions.

Model Selection Guide

Recommended Models by Use Case

Face Consistency Pattern (Nano-Banana Pro)

The Key Insight: For consistent character faces across multiple poses/scenarios, use reference face images with nano-banana-pro:

// Build array of reference image data URIs
function toDataURI(filepath) {
  const buffer = readFileSync(filepath);
  const ext = filepath.endsWith(".png") ? "png" : "jpeg";
  return `data:image/${ext};base64,${buffer.toString("base64")}`;
}

const referenceURIs = referencePhotos.map(p => toDataURI(p));

const response = await replicate.run("google/nano-banana-pro", {
  input: {
    prompt: `${CHARACTER_DESC}, ${scenario.attire}, ${scenario.pose}, ${STYLE}`,
    image_input: referenceURIs,  // Array of data URIs (up to 14 refs)
    aspect_ratio: "1:1",  // Square for avatars
    output_format: "png"
  }
});

Critical: The parameter is image_input (array), not image (single). More references = better face consistency. For flux-2-pro, use input_images instead (up to 8 refs).

Cost Analysis:

• 90 avatar images × $0.025 = $2.25 total (nano-banana)
• vs 90 × $0.04 = $3.60 (Flux 1.1 Pro)
• Savings: 37.5% while maintaining face consistency

Summary

Complete workflow for generating consistent visual character references across multiple scenarios using Flux 1.1 Pro API. Validated through 51 successful image generations across 3 character types.

Critical Insight: Character Consistency Without Reference Images

Common Misconception: Use image_prompt parameter with reference photo for consistency

Reality:

• image_prompt = Flux Redux composition guidance (NOT character consistency)
• Character consistency comes from detailed text descriptions
• Pose variety comes from explicit body position descriptions

Prompt Engineering Pattern

const PROMPT_TEMPLATE = `
${CHARACTER_DESC},          // Detailed physical description
${scenario.attire},         // Scene-appropriate clothing
${scenario.scenario},       // Narrative context
${scenario.pose},          // EXPLICIT body position/gesture
${scenario.environment},   // Setting details
${scenario.lighting},      // Light sources and mood
${scenario.mood},          // Emotional tone
${STYLE_BASE}              // Aesthetic/rendering style
`;

Critical Success Factors

1. Detailed CHARACTER_DESC (creates consistency):

   const CHARACTER_DESC = "Alex Finch, 21, sharp intelligent features, " +
     "short dark brown hair with copper highlights, athletic build, " +
     "confident but approachable demeanor, piercing analytical eyes";
   

2. Explicit Pose Descriptions (creates variety):

   • ❌ Vague: "standing in office"
   • ✅ Specific: "leaning forward with both hands on desk, analyzing data intently"
   • ✅ Specific: "standing with one hand gesturing toward display, explaining concept"

3. Style Consistency:

   // Detective Noir
   const STYLE_BASE = "Professional illustration, high contrast noir lighting, " +
     "dramatic shadows, cinematic composition, sharp detail, moody atmosphere";
   
   // Fantasy Wonderland
   const STYLE_BASE = "Soft fantasy illustration, pastel color harmony, " +
     "ethereal glow, magical atmosphere, dreamy composition";
   

Character-Agnostic Template Structure

// Configuration Block (customize per character)
const CHARACTER = {
  name: "Character Name",
  description: "detailed physical appearance",
  age: number
};

const STYLE = "aesthetic appropriate to narrative";

const SCENARIOS = [
  {
    id: "001",
    title: "Scene Title",
    scenario: "narrative context",
    attire: "specific clothing details",
    pose: "EXPLICIT body position, hands, gaze",
    environment: "setting details",
    lighting: "light sources and mood",
    mood: "emotional tone"
  }
  // ... 16 more scenarios
];

// Generation Engine (reusable)
async function generateScene(scenario) {
  const prompt = buildPrompt(CHARACTER, scenario, STYLE);
  return await replicate.run("black-forest-labs/flux-1.1-pro", {
    input: {
      prompt,
      aspect_ratio: "3:4",
      output_format: "png",
      output_quality: 100,
      safety_tolerance: 2
    }
  });
}

Validated Results

Successful Generations (Feb 2026):

• Alex: 17/17 professional noir scenes
• Iris: 17/17 wonderland magic scenes
• Maya: 17/17 teen life scenes

Quality Metrics:

• 100% generation success rate (with retry handling)
• 17 unique poses per character
• Visual consistency maintained across all images
• Style consistency within each aesthetic type

Troubleshooting

Safety Filter False Positives

Issue: Child character poses sometimes trigger safety filters

Example: "sitting on ground, knees drawn up" (blocked)

Solution: Adjust to neutral poses: "sitting cross-legged leaning forward"

Pattern: Avoid poses that could be misinterpreted for child characters

Rate Limiting

Issue: 429 Too Many Requests from Replicate API

Solution: Exponential backoff retry pattern

async function retryWithBackoff(fn, maxRetries = 3) {
  for (let i = 0; i < maxRetries; i++) {
    try {
      return await fn();
    } catch (error) {
      if (error.response?.status === 429 && i < maxRetries - 1) {
        const delay = 2000 * Math.pow(2, i);
        await new Promise(r => setTimeout(r, delay));
        continue;
      }
      throw error;
    }
  }
}

Pose Repetition

Symptom: All images show similar body positions

Diagnosis: Pose descriptions too vague

Fix: Be cinematically specific

• ✅ "leaning against doorframe, arms crossed, skeptical expression"
• ✅ "crouched examining ground, one hand touching surface"
• ✅ "mid-stride walking forward, confident purposeful movement"

Cost and Performance

Default: Nano-Banana Pro (Face Consistency)

• Model: google/nano-banana-pro (recommended)
• Cost: $0.025 per image
• Generation Time: ~15-30 seconds per image
• Best For: Character avatars, portraits, face consistency
• Requires: Reference face image

Fast Alternative: Nano-Banana 2 (Same API, Faster)

• Model: google/nano-banana-2 (Gemini 3.1 Flash Image)
• Cost: $0.067 per 1K output image ($0.101/2K, $0.151/4K)
• Generation Time: Faster than nano-banana-pro
• API: Identical image_input parameter (array, up to 14 refs)
• Best For: High-volume generation, rapid iteration

Scene Model: Flux 1.1 Pro (No Reference)

• Model: black-forest-labs/flux-1.1-pro
• Cost: $0.04 per image
• Generation Time: ~30-60 seconds per image
• Best For: Full scenes, no reference image needed
• Aspect Ratio: 3:4 portrait recommended
• Output Format: PNG at quality 100 for archival

Economic Comparison:

Note: nano-banana-2's per-image cost at 1K is higher than nano-banana-pro, but it's significantly faster — trade cost for speed when doing rapid iteration.

Cross-Project Applicability

Use Cases:

• Book character reference sheets
• Visual novel character consistency
• Game concept art
• Marketing material uniformity
• Story illustration preparation

Character Types Validated:

• Young adult detective (noir aesthetic)
• Teen fantasy character (ethereal aesthetic)
• Contemporary teenager (realistic aesthetic)

Implementation

Dependencies:

• Replicate API account and token
• Node.js with ES modules support

File Organization:

characters/
  {character-slug}/
    images/
      {collection-name}/
        001-{scenario-title}.png
        002-{scenario-title}.png
        generation-report.json

Package.json script:

"scripts": {
  "generate:character": "node scripts/generate-character-reference.js"
}

Confidence Level

High — Empirically validated through:

• 51 successful image generations
• 3 different character types
• 3 different aesthetic styles
• 100% success rate (with retry handling)