How to Use the GPT-Image2 Prompt System for Batch Generation

The GPT-Image2 prompt system enables scalable batch generation through JSON prompt templates, the centralized data/style-library.json catalog, and the gpt-image-2-style-library CLI that handles async polling and manifest generation for hundreds of images.

The freestylefly/awesome-gpt-image-2 repository implements a Prompt-as-Code engine designed for programmatic image generation at scale. Mastering the GPT-Image2 prompt system for batch generation allows you to convert single structured prompts into cohesive visual series—such as product catalogs or UI mockup sets—while maintaining strict style consistency across hundreds of assets.

Core Architecture Components

The repository organizes its batch generation capabilities into three interconnected layers that separate template management from execution.

Prompt Templates Layer

The foundation of batch generation relies on JSON prompt templates stored in docs/templates.md. This file contains a library of 21 industrial-grade templates spanning visual domains like UI screenshots, infographics, posters, and product photography. JSON templates are essential for batch workflows because they provide validated, structured skeletons that accept programmatic variable injection without breaking prompt syntax. Each template includes "avoid-pitfalls" guides that prevent common generation errors when scaling to large datasets.

Style Library and Agent Skill

Consistency across batch items depends on the centralized style catalog. The repository shares a single JSON catalog between the website and CLI at data/style-library.json, enabling programmatic style selection for any template. The Agent Skill defined in agents/skills/gpt-image-2-style-library/SKILL.md bridges this catalog with generation backends, allowing both automated scripts and conversational agents to resolve style tags (e.g., Minimalist, Neon, Cyberpunk) into valid API parameters.

Generation Backend

The execution layer communicates with supported providers such as APIMart or HiAPI through an async API abstraction. As documented in the Website Auth & Generation section of README.md, the backend authenticates via environment variables like APIMART_API_KEY and supports batch submissions with polling. This architecture prevents timeouts during large jobs by returning a task_id immediately and polling for completion status, enabling you to launch hundreds of concurrent generation tasks without blocking your workflow.

Step-by-Step Batch Generation Workflow

The repository implements a four-phase workflow for converting templates into image series, demonstrated concretely in Case 166 ("Twelve Gold Saints Card Set") at line 404 of README.md—a unified-style card series generated as a batch.

1. Select a JSON template from docs/templates.md. For batch operations, JSON templates are preferred over plain-text because they maintain structural stability when edited programmatically.

2. Construct a parameter array where each object represents one image in the batch. Vary business-specific fields (product names, colors, text content) while keeping structural fields constant.

3. Iterate and dispatch using the gpt-image-2-style-library CLI or direct API calls. The CLI automatically injects the required task_id, manages async polling, and writes results to disk, eliminating manual coordination for each image.

4. Collect results from the generated JSON manifest files (*.json). Each manifest contains the final image URL, metadata, and generation parameters, which you can feed into downstream pipelines for gallery assembly or PDF generation.

Critical Consistency Tip: Always lock the type, layout, and style fields in your JSON template before varying business-specific parameters. This prevents the model from hallucinating unrelated visual elements and ensures your batch output maintains a unified design language.

Practical Implementation Examples

Batch Generation with Node.js and CLI

The following script reads a UI template, merges batch-specific variables, and delegates execution to the CLI:

// batch-ui.js – run with `node batch-ui.js`
const fs = require('fs');
const { execSync } = require('child_process');

// 1️⃣ Load the JSON template (the same one shown in docs/templates.md)
const template = JSON.parse(fs.readFileSync('templates/ui-screenshot.json', 'utf8'));

// 2️⃣ Define batch parameters (could also be read from CSV)
const batch = [
  { product: 'Fitness App', primary_color: 'Neon Green', typography: 'Clean sans-serif' },
  { product: 'Travel Planner', primary_color: 'Sky Blue', typography: 'Rounded' },
  { product: 'Finance Dashboard', primary_color: 'Gold', typography: 'Serif' },
];

// 3️⃣ Iterate and generate
batch.forEach((item, idx) => {
  const prompt = {
    ...template,
    product: item.product,
    style: { ...template.style, primary_color: item.primary_color, typography: item.typography },
  };
  // Write a temporary prompt file
  const promptPath = `tmp/prompt-${idx}.json`;
  fs.writeFileSync(promptPath, JSON.stringify(prompt, null, 2));

  // 4️⃣ Call the CLI (the CLI reads the JSON and sends it to the provider)
  console.log(`🖼️ Generating image ${idx + 1}/${batch.length} – ${item.product}`);
  execSync(`npx gpt-image-2-style-library generate ${promptPath} --output results-${idx}.json`, { stdio: 'inherit' });
});

This approach leverages the CLI's built-in knowledge of data/style-library.json for style resolution and handles provider authentication via your environment configuration.

Using the Agent Skill for Automated Series

For non-developers or rapid prototyping, install the repository's skill into Claude-Code:

/plugin marketplace add freestylefly/awesome-gpt-image-2
/plugin install gpt-image-2-style-library@awesome-gpt-image-2

Then execute batch generation through natural language commands:

Use gpt-image-2-style-library to generate a 5-card product series for "Eco-Smart Water Bottle".

The skill automatically selects the product template, fills series placeholders, and returns a JSON array of image URLs—performing batch generation without requiring custom code.

Direct API Integration for Custom Pipelines

For advanced pipelines requiring custom retry logic or integration with existing infrastructure, call the provider API directly:

curl -X POST https://api.apimart.ai/v1/generate \
  -H "Authorization: Bearer $APIMART_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
        "model": "gpt-image-2-sunburst",
        "prompt": {
          "type":"Product",
          "product":"Eco-Smart Water Bottle",
          "layout":"Hero shot",
          "style":{"theme":"Minimalist","primary_color":"Aqua","typography":"Modern"},
          "constraints":"high-resolution, 9:16"
        }
      }' | jq '.data.url'

Wrap this call in a loop over your parameter array to process lists of products. Store the returned URLs in your batch manifest for downstream processing.

Summary

• JSON templates from docs/templates.md provide the structural foundation for reproducible batch generation.
• The data/style-library.json catalog ensures consistent styling across all items in a batch when accessed via the gpt-image-2-style-library CLI or skill.
• Async polling built into the CLI prevents timeouts when submitting hundreds of generation tasks concurrently.
• Locking structural fields (type, layout, style) while varying content fields prevents visual drift in large batches.
• The repository includes a concrete reference implementation in Case 166 (Twelve Gold Saints Card Set) demonstrating unified series generation.

Frequently Asked Questions

What template format should I use for batch generation?

Use JSON templates from docs/templates.md. According to the repository source code, JSON templates are explicitly recommended for batch work because they separate structure from content, making them easy to edit programmatically while maintaining prompt integrity across hundreds of generations.

How does the CLI handle authentication with image providers?

The gpt-image-2-style-library CLI reads authentication credentials from environment variables such as APIMART_API_KEY, as defined in the Website Auth & Generation section of README.md. It automatically injects these credentials into API requests, so your batch scripts never need to handle raw API keys.

Can I generate image batches without writing custom scripts?

Yes. By installing the Agent Skill from agents/skills/gpt-image-2-style-library/SKILL.md into Claude-Code, you can execute batch generation using natural language commands. The skill handles template selection, parameter variation, and result aggregation automatically.

How do I maintain visual consistency across hundreds of generated images?

Lock the type, layout, and style fields in your JSON template to fixed values, then vary only business-specific fields like product names or taglines. This constraint prevents the underlying model from introducing unexpected visual elements and ensures your batch output functions as a cohesive series.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}