I wanted to generate images from CFML. Not call out to some heavyweight Python service, not stand up a GPU — just write a prompt in Lucee and get a picture back. So I built a small library that wraps Cloudflare’s Workers AI image models, and the whole thing comes down to this:

cf  = new cloudflareimages.CloudflareImages();   // creds from env
img = cf.generate( prompt = "a lighthouse on a rocky coast at sunset, painterly" );
img.toFile( expandPath( "./lighthouse.jpg" ) );

That call produced this:

Try it right now

I put a public demo up — type a prompt, hit generate, get an image:

cfml-image-with-cloudflare.kukiel.dev

It’s capped at a handful of images a day so it stays free, but it’s a real, live CFML app calling real AI models. Here are a few things people (well, me) have made with it:

The annoying part it hides

Cloudflare gives you a bunch of text-to-image models, and they don’t agree on how to answer. Most of the Stable Diffusion family hand you raw image bytes. Flux hands you JSON with the image base64-encoded inside it. If you call the API directly you end up writing two code paths and a pile of content-type sniffing.

The library makes that disappear. Whatever model you pick, you get back the same GenerationResult:

img.getBinary();    // the raw bytes
img.toBase64();     // base64 (handy for a data: URI)
img.toFile( path ); // write it to disk

The ugly part lives in exactly one place — a response normalizer — so you never think about it. Swap @cf/black-forest-labs/flux-1-schnell for @cf/stabilityai/stable-diffusion-xl-base-1.0 and your code doesn’t change.

More than just generate

It also does image-to-image and inpainting, and it can list the available models:

// reimagine an existing picture
img = cf.imageToImage( prompt = "make it winter", image = expandPath("./summer.png") );

// paint something into a masked region
img = cf.inpaint( prompt = "a hat", image = photo, mask = maskBytes );

// what can I use?
models = cf.listModels();

When something goes wrong it throws typed exceptions, so you can actually react to what failed — a missing token (ConfigError), Cloudflare saying no (APIError, with its real message and code), a network blip (TransportError). The demo uses that to show a friendly “that prompt was blocked” message when the safety filter trips, instead of a raw stack trace.

Testing without burning money

The bit I’m happiest with: the test suite never calls Cloudflare. The one place that touches the network — a tiny Transport component — is swappable, so the tests inject a fake that returns canned responses (raw bytes for SDXL, base64 JSON for Flux). The whole suite runs offline, in milliseconds, costs nothing, and still proves the SDXL-vs-Flux normalization actually works. There’s one live smoke test that only wakes up if real credentials are present.

flowchart: your code → facade → (validate) → Transport (the only cfhttp)
           → Cloudflare → normalizer → GenerationResult → back to you

Each piece is one small CFC with one job. That’s what makes it a drop-in: copy the folder, point it at a Cloudflare account, ask for a picture.

Getting a key

You need a free Cloudflare account and a Workers AI API token (dashboard → AI → Workers AI → REST API). The free tier gives you 10,000 “Neurons” a day at no charge — enough for a few hundred images depending on size. A 512×512 image is cheap; 1024×1024 costs about 4× more, so the demo runs at 512 to stretch the free allowance.

Grab the code

It’s all open source — cloudflareimages on GitHub. Clone it, copy the cloudflareimages/ folder into your app, and you’re away. The README has the full API, a Mermaid diagram of how the pieces fit, and the TestBox suite. Use it, fork it, send a PR, and please share it around — the CFML world could use more of this.

If you give the demo a spin, let me know what you make.