# Eduwill Banner

> Use when creating banner images for Eduwill (에듀윌) or similar Korean education companies. Triggers on keywords like 에듀윌, 배너, banner, 교육 배너, 공인중개사, 합격, eduwill. Uses Replicate Nano Banana Pro (google/nano-banana-pro) with reference images for high-quality AI-generated marketing banners. Also applies when user provides reference banner images and wants similar style banners generated.

- Skill: `chacha95/eduwill-banner` (Agent Skill)
- Install (CLI): `npx skillmds@latest add chacha95/eduwill-banner`
- Raw SKILL.md: https://api.skillmd.com/api/skills/chacha95/eduwill-banner/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Marketing & Growth
- Author: chacha95 (https://skillmd.com/u/chacha95)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/chacha95/eduwill-banner

---


# Eduwill Banner Generator

## Overview

Generate high-quality Korean education marketing banners using **Replicate Nano Banana Pro** (`google/nano-banana-pro`) with reference image style transfer. The key to quality output is **uploading reference images via `replicate.files.create()`** and passing the URL through `image_input` — not base64, not prompt-only.

## When to Use

- User asks to create 에듀윌/Eduwill banners
- User provides reference banner images and wants similar style
- Korean education company marketing banner generation
- Any banner task mentioning 공인중개사, 합격, 교육 플랫폼

**When NOT to use:** Simple text-only banners without reference images (use satori/sharp instead).

## Brand Reference

```yaml
eduwill_colors:
  yellow: "#FCC300" # Primary brand color
  navy: "#111E50" # Trust/premium
  sky_blue: "#55ACE2" # Honesty
  red: "#F24646" # Passion
  cta_red: "#F63B28" # Web CTA buttons
  text: "#222222" # Body text

eduwill_fonts:
  primary: "Noto Sans KR" # Open substitute for 에듀윌 합격체
  web_ui: "Spoqa Han Sans Neo"
```

## Core Workflow

```dot
digraph banner {
  "Reference image exists?" [shape=diamond];
  "Upload via replicate.files.create()" [shape=box];
  "Create prediction with image_input" [shape=box];
  "Poll until succeeded" [shape=box];
  "Download result" [shape=box];
  "Optional: composite text with sharp/canvas/pillow" [shape=box];

  "Reference image exists?" -> "Upload via replicate.files.create()" [label="yes"];
  "Reference image exists?" -> "Prompt-only generation" [label="no (lower quality)"];
  "Upload via replicate.files.create()" -> "Create prediction with image_input";
  "Create prediction with image_input" -> "Poll until succeeded";
  "Poll until succeeded" -> "Download result";
  "Download result" -> "Optional: composite text with sharp/canvas/pillow";
}
```

## Quick Reference

| Parameter              | Value                    | Notes                                                                                        |
| ---------------------- | ------------------------ | -------------------------------------------------------------------------------------------- |
| Model ID               | `google/nano-banana-pro` | NOT `nanonobandana/...`                                                                      |
| `aspect_ratio`         | `"16:9"`                 | Standard banner. Default is `"match_input_image"` when `image_input` provided                |
| `resolution`           | `"2K"`                   | High quality                                                                                 |
| `output_format`        | `"png"`                  | Lossless                                                                                     |
| `image_input`          | `[uploaded_url]`         | Must be real URL from `replicate.files.create()`, not base64. Supports up to 14 images       |
| `safety_filter_level`  | `"block_only_high"`      | Most permissive. Options: `block_low_and_above`, `block_medium_and_above`, `block_only_high` |
| `allow_fallback_model` | `false`                  | Falls back to bytedance/seedream-5 if true and model is at capacity                          |

## Implementation

### Step 1: Upload Reference Image (Critical)

```javascript
import Replicate from "replicate";
import fs from "fs";
import "dotenv/config";

const replicate = new Replicate({ auth: process.env.REPLICATE_API_TOKEN });

// MUST upload to get a real URL — base64 data URIs don't work reliably
const buf = fs.readFileSync("inputs/reference-banner.png");
const file = await replicate.files.create(buf, {
  filename: "reference.png",
  content_type: "image/png",
});
const fileUrl = file.urls?.get || file.url;
```

### Step 2: Generate with Polling (High Demand Resilience)

```javascript
// Use predictions API with polling — replicate.run() fails under high demand
const pred = await replicate.predictions.create({
  model: "google/nano-banana-pro",
  input: {
    prompt:
      "Recreate this Korean education banner in the same style and layout. [specific details about what to keep/change]",
    image_input: [fileUrl],
    aspect_ratio: "16:9",
    resolution: "2K",
    output_format: "png",
    safety_filter_level: "block_only_high",
  },
});

// Poll until complete
let p = pred;
while (
  p.status !== "succeeded" &&
  p.status !== "failed" &&
  p.status !== "canceled"
) {
  await new Promise((r) => setTimeout(r, 5000));
  p = await replicate.predictions.get(pred.id);
}

if (p.status === "succeeded") {
  const url = typeof p.output === "string" ? p.output : p.output?.[0];
  const resp = await fetch(url);
  fs.writeFileSync("output/banner.png", Buffer.from(await resp.arrayBuffer()));
}
```

### Step 3: Retry on High Demand (E003)

```javascript
// Nano Banana Pro frequently returns E003 under load — always retry
async function generateWithRetry(input, output, retries = 5) {
  for (let i = 0; i < retries; i++) {
    try {
      const pred = await replicate.predictions.create({
        model: "google/nano-banana-pro",
        input,
      });
      let p = pred;
      while (!["succeeded", "failed", "canceled"].includes(p.status)) {
        await new Promise((r) => setTimeout(r, 5000));
        p = await replicate.predictions.get(pred.id);
      }
      if (p.status === "succeeded") {
        const url = typeof p.output === "string" ? p.output : p.output?.[0];
        const resp = await fetch(url);
        fs.writeFileSync(output, Buffer.from(await resp.arrayBuffer()));
        return output;
      }
      throw new Error(p.error);
    } catch (e) {
      console.log(`Attempt ${i + 1} failed: ${e.message.substring(0, 60)}`);
      if (i < retries - 1) await new Promise((r) => setTimeout(r, 15000));
    }
  }
  return null;
}
```

## Prompt Engineering for Eduwill Banners

**Always start with:** `"Recreate this Korean education banner in the same style and layout."`

**Then specify what to keep/change:**

| Banner Type       | Prompt Pattern                                                                          |
| ----------------- | --------------------------------------------------------------------------------------- |
| 프로모션 (옐로우) | `"...bright yellow gradient, study book on right, bold text on left, price tag"`        |
| 수상/1위 (골드)   | `"...dark gold gradient, sparkle confetti, golden trophy, award celebration"`           |
| 강사진 (베이지)   | `"...cream beige background, professional instructors in suits, CTA button"`            |
| 설명회 (다크)     | `"...dark charcoal background, LIVE badge, instructors on right, yellow CTA"`           |
| 캐릭터 (화이트)   | `"...white background, 3D cartoon students, speech bubbles, yellow bar at bottom"`      |
| 인포그래픽        | `"...split layout, comparison cards, icons (brain/lightbulb/clock), clean infographic"` |

## Post-Processing with Other Methods

After AI generation, composite text overlays using:

- **sharp**: SVG text overlay + resize → `sharp(aiImage).composite([{input: svgBuffer}]).toFile()`
- **canvas**: `loadImage(aiImage)` → `ctx.drawImage()` → draw text/shapes on top
- **pillow**: `Image.open(aiImage)` → `ImageDraw` text + gradient overlay
- **html-to-image**: AI image as `background:url(data:image/png;base64,...)` in HTML template
- **satori**: AI image as `<img>` in JSX tree, text elements overlaid

## Common Mistakes

| Mistake                            | Fix                                                                    |
| ---------------------------------- | ---------------------------------------------------------------------- |
| Using `replicate.run()`            | Use `predictions.create()` + polling — more resilient to E003          |
| Passing base64 to `image_input`    | Upload via `replicate.files.create()` first, pass real URL             |
| Wrong model ID `nanonobandana/...` | Correct: `google/nano-banana-pro`                                      |
| Prompt-only without reference      | Always provide `image_input` with reference for quality                |
| No retry logic                     | E003 (high demand) is frequent — always retry 3-5 times with 15s delay |
| Using `1K` resolution              | Use `2K` for marketing-quality banners                                 |

## Prerequisites

```bash
npm install replicate dotenv
# .env must contain REPLICATE_API_TOKEN=r8_xxxxx
```

## Project Structure

```
banner-generator/
├── .env                  # REPLICATE_API_TOKEN
├── inputs/               # Reference banner images
├── output/               # Generated banners
├── fonts/
│   ├── NotoSansKR-Bold.ttf
│   └── NotoSansKR-Regular.ttf
└── src/
    └── replicate/
        └── generate.js   # Core generation module
```