Why AI Images Suck — And the Pipeline We Built to Fix It

December 12, 2025

Why AI Images Suck — And the Pipeline We Built to Fix It

We manage over $1M/month in paid media for a digital publisher with more than 1,000 promotable pieces of content. Our team? Just five media buyers. No creative team. No creative budget. Compensation is base retainer plus profit share — if we scale, we win. If we don’t, we don’t eat. At first, we hacked it together with Canva and Midjourney. Manual. Slow. Inconsistent. A single ad variant could take 10–15 minutes to produce.

​

Scale was the problem.

​

So we built LaunchBox, our internal platform for creative generation, campaign building, and performance ops. On the image side, it used a cheap, fast model (Flux Schnell) that could churn out thousands of visuals for a few bucks. We layered static ad templates over those images and fed offer context to an LLM to write the text. We were churning out hundreds of creative variations, but there was a problem…

​

Our best-performing ads were still old ones we made manually in Canva.

​

We needed a pipeline that actually delivered creative as good as what we made manually.

Our Solution: A 3-Step Automated Pipeline

Here’s the framework we ended up with:

  1. Sync content from clients and publishers into a central database
  2. Use content context to generate images in bulk through structured LLM processing
  3. Use images + context to generate banners automatically with matched copy and overlays

Step 1: Content Sync

Every image starts with real content — a quiz question, product, landing page, or offer. That content flows into our central database with structured metadata. It sounds simple, but it’s the foundation the rest of the pipeline depends on.

The real challenge? Getting that content reliably. Most clients don’t have APIs, and modern websites use JavaScript rendering, lazy loading, and dynamic content. We needed a scraper that could handle all of that without breaking. Read the full technical breakdown of how we built it.

Example Records:

Quiz: Country Classics: Can You Name These Legendary Country Music Stars?

Question: Which female artist became a country icon with hits like ‘Delta Dawn’ and ‘Two Sparrows in a Hurricane’?

Answers: Lee Ann Womack, Wynonna Judd, Patty Loveless

Correct Answer: Tanya Tucker

Explanation: Tanya Tucker burst onto the country scene as a teenager with her smash hit ‘Delta Dawn’ in 1972 and…

​

Tip

Implementation note: If your content lives in a CMS with API access, syncing is trivial. But even for clients without an API, we’ve made this dead simple. We use ScrapingFish, a $0.002/request scraping API via a tiny edge function to pull structured data from public URLs. No brittle scrapers, no wasted time, no maintenance overhead. Just a clean content ingestion layer we can build on. Learn how we built our content scraper.

Step 2A: Structured Prompt Generation

Text-to-image models like Flux aren’t ChatGPT. You can’t just dump content and say “Make me an interesting image about XYZ.” The prompts need to be literal, concrete, and visual. They need specific subjects, composition details, lighting, and style direction—not abstract concepts or conversational requests.

The solution: You need a small LLM function to build the prompt. We run every input through a custom LLM template first. We use gpt-4o-mini with structured system prompts that are tailored by content type: quiz questions, personality results, DTC offers, editorial content.

What we settled on:

We built two core templates that handle the bulk of our content:

1. Quiz Questions Template

System Prompt:
You are an expert at creating visual prompts for AI image generation (specifically for Flux model).
Your task: Convert quiz questions into descriptive, visual prompts that will generate educational, engaging images.

Guidelines:
- ALWAYS use specific names, brands, and subjects from the question (e.g., "Muhammad Ali", "Squishmallow", "iPhone")
- Create vivid, concrete visual descriptions (not abstract concepts)
- For trivia questions: Include the subject's name, use explanation for context and details
- For personality questions: Create scenes that represent the scenario
- Extract key details from the correct answer and explanation
- Avoid generic descriptions - be specific about WHO or WHAT
- Avoid text, words, or questions in the image
- Use photography/illustration style descriptions
- Be specific about style, lighting, composition, mood
- Keep prompts under 200 characters

Output ONLY the image prompt, nothing else.

2. Offer Content Template

System Prompt:
You are an expert at creating visual prompts for AI image generation (specifically for Flux model).
Your task: Convert content paragraphs into descriptive, visual prompts that will generate engaging images.

Guidelines:
- Use specific details from the title and content
- Create vivid, concrete visual descriptions (not abstract concepts)
- Extract key subjects, places, actions, or themes
- Avoid text, words, or paragraphs in the image
- Use photography/illustration style descriptions
- Be specific about style, lighting, composition, mood
- Keep prompts under 200 characters

Output ONLY the image prompt, nothing else.

Both templates use:

  • Model: gpt-4o-mini
  • Temperature: 0.7

The output is clean, structured, and immediately usable by the image model.

Example Content → Prompt

Quiz Example:

Tanya Tucker performing on a brightly lit country music stage, wearing a sparkling western outfit and singing with passion. A vintage band and guitars fill the background as a lively crowd watches. The scene is bold, energetic, and iconic.

Article Example:

A close-up of a young woman’s startled face, wide eyes and pale expression, with a blurred background. A man stands in front, anxious, under soft, dramatic lighting, capturing intense emotion.

Step 2B: Fast, Scalable Generation (Throughput Beats Perfection)

We don’t chase the “best” model. We use the fastest model that’s good enough. You need volume to test effectively—throughput beats polish. You can’t learn what converts if you can’t generate enough variants to actually test.

Our current model: prunaai/flux-fast at ~$5 per 1,000 images. Solid enough for ads, scroll-stoppers, and creative variants.

Tip

For celebrity likenesses or highly detailed character work, we fall back to Midjourney. Some use cases still need that extra fidelity. But for 90% of our volume, fast and cheap wins.

Example Content → Prompt → Image

Quiz Example:

Quiz example image

Article Example:

Content example image

​

Step 3: Automated Banner Assembly (From Raw Images to Campaign-Ready Assets)

Raw images are just the start. The real leverage comes from turning them into finished, campaign-ready assets automatically.

Every image is fully versioned and traceable—content source, LLM prompt, model parameters, campaign association, timestamps. This metadata makes automation possible.

Because we know exactly where each image came from and what it’s for, we can automatically match images to branded ad templates, pull headlines and CTAs from source content, generate multiple banner sizes in one pass, and maintain full version control with on-demand regeneration.

We use Templated.io to turn raw images into static ads with overlays, headlines, and CTAs. The real advantage is contextual matching: image, copy, and offer all come from the same source. The system knows which headline goes with which visual because they were generated from the same content input.

No more spreadsheets tracking “which image goes with which headline.” No more manual assembly. The pipeline handles it.

Quiz Example: Pipeline Progression

Quiz creative example

Article Example: Pipeline Progression

Content creative example

​

Why It Works: Facebook’s Andromeda Update

Facebook’s Andromeda update changed the game for us. The algorithm now rewards creative freshness, variation, and contextual fit across different delivery buckets. We explore variations without losing brand coherence because everything comes from structured templates tied to brand guidelines.

This is creative diversity with signal. The algorithm gets what it’s optimized for, and so do we, because we can actually track and attribute what’s working.

Tip

A note on aesthetics: You might be looking at these creatives and thinking, “This doesn’t work for my brand.” But it’s important to understand that we’re doing what works for these offers and this audience. When it comes down to it, what works works. They may not be the prettiest creative, but ugly works.

Results: What This Pipeline Actually Delivers

Here’s what one of those “ugly” ads delivered in the last 2 weeks:

Performance:

  • Spend: $6,062
  • Revenue: $9,751
  • Profit: $3,689
  • ROI: 60.9%
  • CTR: 15.5%
  • Clicks: 77,806
  • Impressions: 502,605

Overall impact: We’ve seen massive increases in our Feed traffic on Meta. 30-day change went from $39K on Facebook Feed to $300K in spend.

See It In Action With Your Own Content

We built this pipeline to scale our own campaigns at Conversion Collective. It solved a real problem we had running high-volume paid media for digital publishers.

Now we’re opening it up for others to test.

If you want to see how contextual image generation works with your actual content, we’ll generate a free creative pack: 25 ad-ready images plus matched copy, built from your real offers or landing pages.

No generic examples. No placeholder content. Just your brand, your offers, and the kind of volume this system makes possible.

👉 Claim your free creative pack here​

How to Build a Content Scraper for Context (Without Breaking)

Monday, Dec 15, 2025

How to Build a Content Scraper for Context (Without Breaking)

Why AI Images Suck — And the Pipeline We Built to Fix It

Friday, Dec 12, 2025

Why AI Images Suck — And the Pipeline We Built to Fix It

See What Our Creative System Can Do for Your Brand

Start with a your free creative audit + ad bundle—no obligation, just real insights and test-ready creative.