Menü
Generate consistent visual character references across multiple scenarios using Flux and nano-banana-pro on Replicate
| type | skill |
| lifecycle | external-dependent |
| inheritance | inheritable |
| name | ai-character-reference-generation |
| description | Generate consistent visual character references across multiple scenarios using Flux and nano-banana-pro on Replicate |
| tier | standard |
| applyTo | **/*character*,**/*portrait*,**/*avatar*,**/*reference*,**/generate-alex* |
| currency | "2026-04-20T00:00:00.000Z" |
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-ultraprovides higher resolution (up to 4MP) with the same API surface.black-forest-labs/flux-2-proadds multi-reference editing.google/nano-banana-2is the faster/cheaper successor to nano-banana-pro (sameimage_inputAPI). Input parameter schema may change between model versions. See also EXTERNAL-API-REGISTRY.md.
| Use Case | Model | Cost | Notes |
|---|---|---|---|
| Face Consistency (Recommended) | google/nano-banana-pro | $0.025/image | Best for maintaining face identity — uses reference image + prompt. Up to 14 refs, 4K output. |
| Face Consistency (Fast/Cheap) | google/nano-banana-2 | $0.067/1K img | Same image_input API as nano-banana-pro. Faster (Gemini 3.1 Flash), similar quality. |
| Detailed Scenes | black-forest-labs/flux-1.1-pro | $0.04/image | High-quality scenes without face reference. Good for backgrounds and environments. |
| Multi-Reference High Quality | black-forest-labs/flux-2-pro | ~$0.05+ | Up to 8 refs via input_images. Text rendering + photorealism. |
| Agent Banners | ideogram-ai/ideogram-v3-turbo | $0.03/image | Best for stylized text + character combinations. Cheapest typography option. |
| High Resolution | black-forest-labs/flux-1.1-pro-ultra | $0.06/image | Up to 4MP output. Use when print-quality needed. |
The Key Insight: For consistent character faces across multiple poses/scenarios, use reference face images with nano-banana-pro:
// Build array of reference image data URIs
function toDataURI(filepath) {
const buffer = readFileSync(filepath);
const ext = filepath.endsWith(".png") ? "png" : "jpeg";
return `data:image/${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), notimage(single). More references = better face consistency. For flux-2-pro, useinput_imagesinstead (up to 8 refs).
Cost Analysis:
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.
Common Misconception: Use image_prompt parameter with reference photo for consistency
Reality:
image_prompt = Flux Redux composition guidance (NOT character consistency)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
`;
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";
Explicit Pose Descriptions (creates variety):
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";
// 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
}
});
}
Successful Generations (Feb 2026):
Quality Metrics:
After generating a character set, use view_image to review outputs without leaving chat:
For batch sets, VS Code's image carousel displays them side-by-side for efficient visual QA.
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
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;
}
}
}
Symptom: All images show similar body positions
Diagnosis: Pose descriptions too vague
Fix: Be cinematically specific
google/nano-banana-pro (recommended)google/nano-banana-2 (Gemini 3.1 Flash Image)image_input parameter (array, up to 14 refs)black-forest-labs/flux-1.1-proEconomic Comparison:
| Character Set | nano-banana-pro | nano-banana-2 | flux-1.1-pro |
|---|---|---|---|
| 17 scenarios | $0.43 | $1.14 (1K) | $0.68 |
| 90 avatars | $2.25 | $6.03 (1K) | $3.60 |
| Full set (100+) | ~$2.50 | ~$6.70 (1K) | ~$4.00 |
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.
Use Cases:
Character Types Validated:
Dependencies:
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"
}
High — Empirically validated through:
Run the 7-step Artificial Critical Thinking pass — Materiality → Hypothesise → Alternatives → Disconfirmers → Audit priors → Severity → Commit-with-marker
Challenge what you think is right — alternative hypotheses, missing data, evidence quality, bias detection, falsifiability, and adversarial review
Step-back protocol — restate, generalise, specialise, invert, ask why, pre-mortem, check stakeholders, and audit framings before solving
Build knowledge bases that build software — research before code, teach before execute
Two-phase brain upgrade — mechanical install via shared core, then LLM-led semantic reconciliation
Create professional ultra-wide cinematic banners for GitHub READMEs using modern AI image models with typography