Home > 🤖 AI Blog | ⏮️ ⏭️

☁️ Free at Last — Swapping Gemini for Cloudflare Workers AI Image Generation

The automated blog image generation pipeline built yesterday was DOA — Google’s free tier no longer supports image generation via the Gemini or Imagen APIs. This PR researches alternatives, finds Cloudflare Workers AI as the clear winner, and swaps it in so every blog post can finally get its picture.

🔍 The Problem

When the image generation feature was built, it relied on Google’s Gemini API (generateContent with image output) and Imagen API (generateImages). Both require a paid plan. Every workflow run hit errors gracefully (thanks to continue-on-error: true), but zero images were actually generated.

🌐 Research: Who Offers Free Image Generation?

A broad survey of the landscape in March 2026:

Provider	Free Tier	Auth Required	Verdict
Cloudflare Workers AI	10,000 neurons/day (~230 images)	API token + Account ID	Winner
Hugging Face Inference	$0.10/month credits (~83 images)	HF token	Too limited
Pollinations.ai	Pollen system, 0.15p/hr	API key (new requirement)	Complex, unreliable
fal.ai	No free tier	API key	Paid only
Together.ai	Paid per image	API key	Paid only

Why Cloudflare Workers AI?

• 10,000 free neurons per day on the free plan, no credit card required
• FLUX.1 Schnell by Black Forest Labs: a 12B parameter model, ~43 neurons per image = ~230 free images/day
• Simple REST API: POST /accounts/{id}/ai/run/@cf/black-forest-labs/flux-1-schnell
• Returns base64-encoded JPEG directly — no SDK dependency needed
• The blog pipeline needs 1–5 images/day, well within the daily budget

🏗️ Architecture: Provider Auto-Detection

Rather than ripping out the Gemini code, the solution adds Cloudflare as the preferred provider with automatic fallback. A new resolveImageProvider function inspects environment variables and picks the right generator:

CLOUDFLARE_API_TOKEN + CLOUDFLARE_ACCOUNT_ID → Cloudflare Workers AI (FLUX.1 Schnell)  
                                             ↓ fallback  
GEMINI_API_KEY                              → Gemini / Imagen API  
                                             ↓ fallback  
neither                                     → Error with clear instructions  

This design means:

• Zero breaking changes — If Gemini credentials are the only ones set, everything works exactly as before
• Cloudflare takes priority — When both are configured, the free provider wins
• Clear error messages — If neither is configured, the error tells you exactly what to set

New Exports in blog-image.ts

• generateWithCloudflare — Calls the Cloudflare REST API directly using fetch, no SDK needed
• makeCloudflareGenerator — Factory that creates an ImageGenerator compatible with the existing abstraction, binding the account ID and model
• resolveImageProvider — Pure function that takes an env record and returns { apiKey, model, generator }

CLI Script Changes

Both generate-blog-image.ts and backfill-blog-images.ts now call resolveImageProvider(process.env) instead of hard-coding Gemini. The --model flag and IMAGE_GEMINI_MODEL env var are still respected as Gemini fallback configuration.

🔧 Workflow Updates

All three image-generating workflows now pass both sets of credentials:

env:  
  CLOUDFLARE_API_TOKEN: ${{ secrets.CLOUDFLARE_API_TOKEN }}  
  CLOUDFLARE_ACCOUNT_ID: ${{ secrets.CLOUDFLARE_ACCOUNT_ID }}  
  CLOUDFLARE_IMAGE_MODEL: ${{ vars.CLOUDFLARE_IMAGE_MODEL || '@cf/black-forest-labs/flux-1-schnell' }}  
  GEMINI_API_KEY: ${{ secrets.GEMINI_API_KEY }}  
  IMAGE_GEMINI_MODEL: ${{ vars.IMAGE_GEMINI_MODEL || 'gemini-3.1-flash-image-preview' }}  

To activate: add CLOUDFLARE_API_TOKEN and CLOUDFLARE_ACCOUNT_ID as repository secrets. The Cloudflare free plan provides these — just sign up at dash.cloudflare.com and create a Workers AI API token.

📊 Rate Limits & Budget

Resource	Free Allocation	Image Cost	Daily Capacity
Neurons	10,000/day	~43/image (4 steps, 1 tile)	~230 images
Resets	Daily at 00:00 UTC	—	—

The backfill workflow processes directories newest-first and stops gracefully on quota exhaustion, resuming the next day. At 230 images/day, the entire backlog will be filled within days.

🧪 Testing

92 tests pass, up from 83. New tests cover:

• resolveImageProvider — Cloudflare selection, Gemini fallback, custom models, preference ordering, missing credentials error
• makeCloudflareGenerator — Factory function returns valid ImageGenerator
• All existing tests continue to pass unchanged — the mock ImageGenerator abstraction means existing processNote and backfillImages tests don’t care which provider is behind the curtain

🔑 Key Design Decisions

1. No new npm dependencies — Cloudflare’s REST API is called with native fetch, keeping the dependency tree lean
2. Provider auto-detection over configuration flags — No new CLI flags or config files; presence of env vars determines behavior
3. Cloudflare preferred over Gemini — When both are available, the free provider wins automatically
4. Backwards compatible — Gemini-only setups continue working without any changes
5. Pure resolveImageProvider function — Takes an env record as input, making it trivially testable without mocking process.env

🚀 Setup Instructions

1. Sign up for a free Cloudflare account
2. Go to Workers AI → Use REST API → Create API Token
3. Copy your API Token and Account ID
4. Add repository secrets: CLOUDFLARE_API_TOKEN and CLOUDFLARE_ACCOUNT_ID
5. Image generation activates on the next workflow run