I Built an AI Pipeline That Turns One Idea Into 39 Sticker Packs (Here's the Whole System)

How I went from one prompt to 39 commercially-licensed sticker sets, a Gumroad store,
and 160 indexable pages — without a GPU, on a $7/month VPS.

Last month I had a problem most creators will recognise: I could design one good sticker
pack, but scaling to a full catalogue meant weeks of manual work I didn't have time for.

So I did the boring, engineer-brained thing. I stopped designing stickers and started designing
a pipeline.

Today that pipeline has produced 39 sticker themes and 160 pieces of
indexable content. Everything runs on a single 1-vCPU VPS with no GPU. Here's exactly how it
works, including the parts that went wrong.

The stack (nothing exotic)

• A tiny Linux VPS (1 vCPU, 2 GB RAM — the cheap tier)
• Python for orchestration
• A text-to-image model for the raw art
• A vector tracer to convert raster art into clean EPS/SVG
• The Gumroad API for the storefront

No fine-tuning. No custom model. No GPU. The whole thing costs less than a coffee per week to run.

Step 1: Generate, but with a strict spec

The temptation is to prompt "cute stickers" and hope. That produces garbage you can't sell.

Instead I locked the spec before generating anything:

• One consistent art style across every icon (same line weight, same palette)
• Fixed canvas sizes so every file is print-ready
• Transparent backgrounds, always
• A hard cap on file size so nothing gets rejected on upload

The output is a raster sheet: six icons on one transparent canvas. Consistent, but not yet sellable.

Step 2: Vectorise (this is the part people skip)

Marketplaces and Cricut users want vectors. A PNG is a picture; an SVG is a file you can cut,
resize and recolour forever.

I run every sheet through a vector tracer, which converts the raster art into paths. The result is
a real EPS + SVG that opens cleanly in Illustrator, Inkscape, Cricut Design Space and Silhouette
Studio.

This one step is the difference between "another PNG pack" and "a professional vector pack".
Buyers notice.

Step 3: Package it so it's usable in 30 seconds

Here's a rule I stole from software: the first 30 seconds decide whether someone uses your product.

So every pack ships as:

• Individual transparent PNGs (drag straight into GoodNotes)
• A single PNG sticker sheet
• An EPS vector (Illustrator 10 compatible)
• An SVG (Cricut / Silhouette)
• A preview image
• A plain-language commercial-use license

No zip full of mystery files. No README that assumes you're a designer.

Step 4: Automate the boring 90%

Uploading, writing descriptions, generating covers, setting categories — this is where creators
burn out. All of it is API calls.

I wrote a small uploader that uploads the file, sets a clean permalink, generates a cover and
thumbnail, writes an SEO-aware description, and publishes the product.

Then I hit one wall that surprised me: Gumroad only lets you create 10 new products per day.
Not a bug — a deliberate anti-spam limit.

The fix was a scheduled script that uploads 10 per day, remembers what it has already done (so it
never double-posts) and goes quiet when there's nothing left to do. Zero babysitting.

Step 5: The traffic problem (the real one)

Here's the uncomfortable truth: Gumroad is a cash register, not a traffic engine. You can have
the best product on the platform and sell nothing if nobody finds you.

So I built a free, automated distribution layer:

• 97 SEO pages on a Google-indexable free host, each targeting a long-tail buyer keyword ("cute dinosaur stickers PNG for Cricut")
• 63 promo pages on an anonymous, indexable paste site
• Nostr posts to an open, account-free social protocol
• A hub page that internally links the whole catalogue together

Every page points back to a free sampler, which points to the paid packs. That's the funnel:
attention → trust → sale.

What actually went wrong

I'll save you the pain:

1. Empty covers killed conversions. Every product launched without a cover image. Covers are the single biggest conversion factor on a marketplace. Fixed by automating cover generation.
2. The API lies by omission. The product list endpoint hides files and covers, so it looked like nothing was attached. The detail endpoint showed everything was fine. Never trust a summary.
3. Free ≠ free to sell. Plenty of "free" clipart is personal-use only. If you sell print-on-demand goods with it, you're exposed. Everything in my packs is explicitly commercial-use licensed.
4. Scaling breaks cheap servers. Running the whole pipeline at once on 2 GB of RAM is how you get a crash at 3am. I now measure peak memory per step and cap concurrency to one process.

The takeaway

The creators winning in 2026 aren't the ones with the best single product. They're the ones who built
a repeatable system — generate, vectorise, package, publish, distribute — and let it run.

You don't need a GPU, a team or a big budget. You need a spec, a pipeline and patience with the boring
parts.

Want to see the output? I put a free pack of cute kawaii stickers (transparent PNG, commercial-use)
here — no email required: Free cute sticker sampler

If you want the full collection, the mega bundle has every theme in EPS + SVG + PNG:
Cute sticker mega bundle

And if you're building something similar, the scripts that do all of the above — generation, the memory guards, the packaging, the publishing — are collected in one place:

The Pipeline Starter Kit — the exact system, not a tutorial.

The playbook version, if you'd rather read the reasoning than run the code: The AI Sticker Factory

More from me, if you liked this:

• How I turn AI images into print-ready vector stickers (no Illustrator, no GPU)
• A Gumroad store with 39 products and no visitors — what I got wrong about distribution