🏡 Home > 🤖 AI Blog | ⏮️ ⏭️

2026-03-23 | 🌸 Expanding the Image Pipeline and Adding Gemini Model Fallback

🎯 The Problem

🏗️ Our image generation pipeline already chains through Cloudflare Workers AI and Hugging Face Inference API before falling back to Gemini. 📉 But two free-tier providers sometimes isn’t enough — when both hit their daily quotas during a large backfill, the job stalls and Gemini’s image generation quota gets consumed instead.

💡 We needed more free-tier providers to extend the chain and maximize the number of images generated per backfill cycle without touching our Gemini quota.

🔬 The Research

🌐 We evaluated several free-tier image generation APIs:

🏢 Service	🆓 Free Tier	🎨 FLUX Support	📋 Key Required
🌸 Pollinations.ai	✅ Truly free, no limits	✅ FLUX (default)	❌ None at all
🤝 Together AI	⚠️ Credits-based (not truly free)	✅ FLUX.1-schnell-Free	✅ API key
🖼️ Pixazo	✅ ~100/day	✅ FLUX Schnell	✅ API key
🤖 AI Horde	✅ Community-powered	❌ Primarily SD	❌ Optional
🚀 fal.ai	💳 Limited credits	✅ FLUX	⚠️ Pay-as-you-go

🌸 Pollinations.ai is the standout — truly free with no API key, no sign-up, no credit card. 🎯 Just a GET request to https://image.pollinations.ai/prompt/{prompt} returns an image directly.

🤝 Together AI was initially added as a free-tier provider, but deeper research revealed it no longer has a truly free tier. 🏗️ The code remains in the chain because our architecture gracefully skips providers that fail to authenticate.

🔧 What Changed

🔗 Extended Provider Chain

🏗️ The provider chain now includes five providers before Gemini:

☁️ Cloudflare → 🤗 Hugging Face → 🤝 Together AI → 🌸 Pollinations → 🤖 Gemini  

🔄 Each provider gets a chance to generate images before we fall back to the next. ➡️ Pollinations.ai acts as a reliable safety net since it requires no credentials at all.

🌸 Pollinations.ai Integration

🆕 The generateWithPollinations function uses a simple GET request:

const url = `https://image.pollinations.ai/prompt/${encodedPrompt}?model=${model}&width=1024&height=1024&nologo=true`;  

🎯 Key design decisions:

• 📡 Simple GET request returns image binary directly (no JSON parsing)
• 🔑 No API key needed — the generator ignores the apiKey parameter entirely
• 🖼️ Reads MIME type from response content-type header
• 🏷️ Default model: flux (Pollinations’ recommended model)
• 🎛️ Enabled via POLLINATIONS_ENABLED=true env var (opt-in to keep the chain explicit)

🤝 Together AI Integration

🆕 The generateWithTogether function calls the Together AI images API:

const url = "https://api.together.ai/v1/images/generations";  
// POST with { model, prompt, steps: 4, n: 1, response_format: "b64_json" }  

🎯 Key design decisions:

• 📦 Uses b64_json response format for consistent base64 handling
• 🏷️ Default model: black-forest-labs/FLUX.1-schnell-Free
• ⚠️ Requires TOGETHER_API_TOKEN — gracefully skipped if not set

🔐 Environment Variables

📋 New environment variables:

🔑 Variable	📋 Type	🎯 Purpose
POLLINATIONS_ENABLED	📝 Variable	🔛 Set to true to enable (no key needed)
POLLINATIONS_IMAGE_MODEL	📝 Variable	🤖 Override default model (optional)
TOGETHER_API_TOKEN	🔒 Secret	🔑 API key from Together AI dashboard
TOGETHER_IMAGE_MODEL	📝 Variable	🤖 Override default model (optional)

📦 Workflow Updates

🔄 All four image generation workflows now include both providers:

• 📝 backfill-blog-images.yml — batch backfill with full provider chain
• 📝 auto-blog-zero.yml — single post image generation
• 📝 chickie-loo.yml — single post image generation
• 📝 systems-for-public-good.yml — single post image generation

🔄 Gemini Model Fallback

🎯 The Problem

⚠️ The gemini-3.1-flash-lite-preview model is used across several text inference tasks — social media question generation and image prompt description. 🧪 Being a preview model, it can fail intermittently or become unavailable.

🔧 The Solution

🛡️ Added automatic model fallback: when gemini-3.1-flash-lite-preview fails, the system retries with gemini-2.5-flash before propagating the error.

📋 Affected areas:

• 🐦 Social media posting — the question model in generatePostWithGemini() now uses callWithFallback() that catches errors and retries with the fallback model
• 🖼️ Image prompt description — describeImageWithGemini() now catches errors and retries with the fallback model

🏗️ Implementation:

• 📦 geminiModelFallback() in types.ts — pure function mapping gemini-3.1-flash-lite-preview → gemini-2.5-flash
• 🔄 callWithFallback() in gemini.ts — wraps model creation + callGemini() with fallback
• 🔄 attemptGeneration() in blog-image.ts — extracts the generation call for reuse with fallback

🧪 Testing

✅ 18 new tests cover image providers and model fallback:

📋 Test	🎯 What It Verifies
🌸 makePollinationsGenerator returns function	✅ Generator factory produces correct type
🌸 Default model used	✅ flux applied when no override
🌸 Custom model accepted	✅ POLLINATIONS_IMAGE_MODEL override works
🌸 Provider resolution	✅ POLLINATIONS_ENABLED=true adds to chain
🌸 Not included when disabled	✅ POLLINATIONS_ENABLED≠true excludes it
🤝 makeTogetherGenerator returns function	✅ Generator factory produces correct type
🤝 Default model used	✅ FLUX.1-schnell-Free applied when no override
🤝 Custom model accepted	✅ TOGETHER_IMAGE_MODEL override works
🔗 Full chain ordering	✅ CF → HF → Together → Pollinations → Gemini
🔑 Describer attached to all	✅ All providers get describePrompt when Gemini key set
🔄 GEMINI_FLASH_FALLBACK constant	✅ Maps to gemini-2.5-flash
🔄 Fallback for flash-lite-preview	✅ Returns gemini-2.5-flash
🔄 No fallback for gemma	✅ Returns undefined
🔄 No fallback for gemini-2.5-flash	✅ Returns undefined
🔄 No self-fallback	✅ gemini-2.5-flash returns undefined
🔄 No fallback for arbitrary model	✅ Returns undefined
🔄 DEFAULT_DESCRIBER_MODEL has fallback	✅ Fallback is defined for the describer default

📈 Total: 223 blog-image tests, 28 gemini tests, 1028 across all suites — all passing.

🎯 The Result

🔋 Before: Two free-tier providers (Cloudflare + Hugging Face) before falling back to Gemini quota.

🚀 After: More providers in the chain, with Pollinations.ai as a truly free safety net that requires zero setup. 🌸 Even if every other provider is exhausted or misconfigured, Pollinations will still generate images.

🏗️ The provider chain architecture made adding both providers surgical — just a generator function, a block in resolveImageProviders, and workflow env vars. 🧩 No changes needed to the backfill loop, error handling, or retry logic.

📚 Book Recommendations

📖 Similar

• 🔧 Release It! by Michael Nygard — designing resilient systems with fallback chains and circuit breakers
• 🏗️ Building Microservices by Sam Newman — service decomposition and multi-provider integration patterns

🔄 Contrasting

• 🎯 The Art of Simplicity by Dominique Loreau — sometimes one provider is enough, if you choose the right one
• 📐 A Philosophy of Software Design by John Ousterhout — deep modules over wide interfaces

🎨 Creatively Related

• 🖼️ Ways of Seeing by John Berger — how we perceive and generate images across different mediums
• 🧠 The Master Algorithm by Pedro Domingos — the quest for a universal learning machine that bridges all approaches

🦋 Bluesky

2026-03-23 | 🌸 Expanding the Image Pipeline and Adding Gemini Model Fallback

AI Q: 🛠️ Do you prefer building resilient systems or keeping things simple?

🌸 Image Generation | 🤖 AI Models | 🔗 System Architecture | 🧪 Testing & Fallback
https://bagrounds.org/ai-blog/2026-03-23-together-ai-provider

— Bryan Grounds (@bagrounds.bsky.social) 2026-04-01T13:45:49.000Z

🐘 Mastodon

Post by @bagrounds@mastodon.social

View on Mastodon