Programmatic Image Generation: How to Build a Batch Pipeline That Scales

Most image generation tutorials stop at one image. They show you a curl command, a pretty output, and call it done. That's fine for "generate a cat picture." It's useless when you need 500 images for a real project.

Programmatic image generation — generating images at scale, from code, without human intervention — is a different skill. This guide covers the full pipeline: prompt engineering at scale, batch generation, error handling, async processing, output management, and integration into production systems.

The Three Layers of a Production Image Pipeline

Every production image pipeline has three layers:

Most developers only think about Layer 1. But Layers 2 and 3 are where pipelines live or die.

Layer 1: Prompt Engineering at Scale

When you generate one image, you can lovingly craft the perfect prompt. When you generate 500, you need a prompt system.

The Template Approach

# prompts.py — Centralized prompt templates
from dataclasses import dataclass
from typing import Optional

@dataclass
class ImageJob:
    template: str
    params: dict
    output_path: str
    model: str = "nano-banana-2"
    
PROMPT_TEMPLATES = {
    "product_hero": "E-commerce product photo: {product_name}, {color}, studio lighting, white background, 1024x1024, commercial photography",
    "blog_hero": "Blog header illustration: {topic}, {style} style, {mood} mood, 1200x630, editorial",
    "social_post": "Social media visual: {subject}, {platform} format, {vibe} aesthetic, {dimensions}",
}

def build_prompt(template_key: str, **params) -> str:
    return PROMPT_TEMPLATES[template_key].format(**params)

The Scale-Up Pattern

# Generate 100 product photos from a CSV
import csv, subprocess, json
from concurrent.futures import ThreadPoolExecutor, as_completed

def generate_single(job: ImageJob) -> dict:
    prompt = build_prompt(job.template, **job.params)
    
    result = subprocess.run([
        "anycap", "image", "generate",
        "--prompt", prompt,
        "--model", job.model,
        "--output-format", "json",
        "-o", job.output_path
    ], capture_output=True, text=True)
    
    return {
        "output_path": job.output_path,
        "success": result.returncode == 0,
        "data": json.loads(result.stdout) if result.returncode == 0 else None,
        "error": result.stderr if result.returncode != 0 else None
    }

# Build job list from data
jobs = []
with open("products.csv") as f:
    for row in csv.DictReader(f):
        jobs.append(ImageJob(
            template="product_hero",
            params={"product_name": row["name"], "color": row["color"]},
            output_path=f"output/{row['sku']}.png"
        ))

# Execute with concurrency control
with ThreadPoolExecutor(max_workers=4) as executor:
    futures = {executor.submit(generate_single, job): job for job in jobs}
    for future in as_completed(futures):
        result = future.result()
        status = "✅" if result["success"] else "❌"
        print(f"{status} {result['output_path']}")

Layer 2: Orchestration — The Part Everyone Forgets

Generation is easy. Making it reliable at scale is the real engineering.

Pattern 1: Async Batch Processing

For large batches (100+ images), use async mode to avoid blocking:

# Submit batch job
anycap image generate \
  --prompt "$(python build-prompts.py --csv products.csv)" \
  --model nano-banana-2 \
  --async \
  --batch-size 20 \
  --webhook "https://your-app.com/webhooks/images" \
  -o output/products/

Your webhook receives results as they complete. No polling. No timeout issues.

Pattern 2: Retry with Exponential Backoff

import time, random

def generate_with_retry(job: ImageJob, max_retries: int = 3) -> dict:
    for attempt in range(max_retries):
        result = generate_single(job)
        if result["success"]:
            return result
        
        if attempt < max_retries - 1:
            wait = (2 ** attempt) + random.uniform(0, 1)
            print(f"Retry {attempt + 1}/{max_retries} for {job.output_path} in {wait:.1f}s")
            time.sleep(wait)
    
    return result  # Return the last failure

Pattern 3: Queue-Based Architecture

For production systems, use a proper queue:

# Simple Redis-based job queue
import redis, json

r = redis.Redis()

def enqueue_job(job: ImageJob):
    r.lpush("image_jobs", json.dumps({
        "template": job.template,
        "params": job.params,
        "output_path": job.output_path,
        "model": job.model,
    }))

def worker_loop():
    while True:
        _, job_data = r.brpop("image_jobs")
        job = json.loads(job_data)
        result = generate_single(ImageJob(**job))
        
        if result["success"]:
            r.lpush("image_results", json.dumps(result))
        else:
            r.lpush("image_failures", json.dumps(result))

Layer 3: Integration — Getting Images Where They Need to Go

Upload to S3

import boto3

s3 = boto3.client("s3")

def upload_to_s3(local_path: str, bucket: str, key: str) -> str:
    s3.upload_file(local_path, bucket, key, ExtraArgs={
        "ContentType": "image/png",
        "CacheControl": "public, max-age=31536000",
    })
    return f"https://{bucket}.s3.amazonaws.com/{key}"

Post to CMS

import requests

def update_cms_product_image(sku: str, image_url: str):
    requests.patch(
        f"https://cms.example.com/api/products/{sku}",
        headers={"Authorization": "Bearer $CMS_TOKEN"},
        json={"image_url": image_url}
    )

Notify Your Team

def notify_slack(message: str):
    requests.post(
        "https://hooks.slack.com/services/YOUR/WEBHOOK/URL",
        json={"text": message}
    )

The Complete Pipeline Script

#!/usr/bin/env python3
"""production-pipeline.py — Full image generation pipeline"""

import csv, subprocess, json, time, random, sys
from concurrent.futures import ThreadPoolExecutor, as_completed
from dataclasses import dataclass
import boto3, requests

# --- Configuration ---
S3_BUCKET = "my-assets"
SLACK_WEBHOOK = "https://hooks.slack.com/services/YOUR/WEBHOOK"
MAX_WORKERS = 4
MAX_RETRIES = 3

PROMPTS = {
    "product": "E-commerce photo: {name}, {color}, studio, white bg, 1024x1024",
    "lifestyle": "Lifestyle photo: {name}, {color}, {scene}, natural light, 1024x1024",
}

@dataclass
class Job:
    template: str
    params: dict
    output: str
    model: str = "nano-banana-2"

def generate(job: Job) -> dict:
    prompt = PROMPTS[job.template].format(**job.params)
    for attempt in range(MAX_RETRIES):
        result = subprocess.run([
            "anycap", "image", "generate",
            "--prompt", prompt, "--model", job.model,
            "--output-format", "json", "-o", job.output
        ], capture_output=True, text=True)
        if result.returncode == 0:
            data = json.loads(result.stdout)
            return {"path": job.output, "url": data.get("image_url"), "success": True}
        if attempt < MAX_RETRIES - 1:
            time.sleep((2 ** attempt) + random.uniform(0, 1))
    return {"path": job.output, "success": False, "error": result.stderr}

def upload(path: str) -> str:
    key = path.replace("output/", "")
    s3 = boto3.client("s3")
    s3.upload_file(path, S3_BUCKET, key, ExtraArgs={"ContentType": "image/png"})
    return f"https://{S3_BUCKET}.s3.amazonaws.com/{key}"

def notify(text: str):
    requests.post(SLACK_WEBHOOK, json={"text": text})

def run_pipeline(csv_path: str):
    jobs = []
    with open(csv_path) as f:
        for row in csv.DictReader(f):
            jobs.append(Job("product", {"name": row["name"], "color": row["color"]}, f"output/{row['sku']}.png"))
    
    notify(f"🚀 Pipeline started: {len(jobs)} images")
    
    results = []
    with ThreadPoolExecutor(max_workers=MAX_WORKERS) as executor:
        futures = {executor.submit(generate, job): job for job in jobs}
        for future in as_completed(futures):
            result = future.result()
            if result["success"]:
                result["s3_url"] = upload(result["path"])
                results.append(result)
    
    success = len(results)
    failed = len(jobs) - success
    notify(f"{'✅' if failed == 0 else '⚠️'} Pipeline complete: {success}/{len(jobs)} images. {failed} failed.")
    return results

if __name__ == "__main__":
    run_pipeline(sys.argv[1])

Choosing the Right Model for Your Pipeline

Cost at Scale

Last updated: May 2026.