How we trained our AI to design supplement packaging that actually sells

When we set out to build Productifi, we faced a deceptively hard problem: AI image models can generate beautiful packaging, but most of what they produce won't sell on a Shopify store. There's a gap between aesthetic output and commercial conversion that no off-the-shelf model could close.

This is the story of how we narrowed that gap by training our model on 50,000 bestselling supplement products — and what we learned about what actually drives a customer to click "add to cart."

The dataset that changed everything

We spent six months assembling a labeled corpus of supplement packaging from Amazon, Shopify storefronts, iHerb, and Vitacost. For each product, we recorded not just the design but the metadata that mattered: monthly revenue, conversion rate, return rate, average review score, and category positioning.

The pattern that emerged surprised us. The bestsellers were not the most beautiful products. They were the most legible — high contrast typography, clear ingredient hierarchy, immediate category signaling. A premium ashwagandha bottle that looked like it belonged in Whole Foods consistently outsold the same formula in a generic clinical bottle, even when both used identical ingredients.

Three things that didn't work

Before we landed on the current architecture, we hit a few dead ends worth sharing:

• Style transfer from luxury cosmetics. We thought borrowing visual language from Glossier and Aesop would translate. It didn't — supplement buyers want to see what the product does, not how it makes them feel.
• Template-based generation. Filling in pre-designed templates with AI-generated copy felt cheap and converted poorly. Customers can spot a template instantly.
• Pure text-to-image generation. Stable Diffusion and DALL-E produce stunning bottles but the labels are gibberish, the dosages are made up, and the FDA-required panels are missing.

The breakthrough: structured generation

Our current model doesn't generate packaging as a single image. It generates a structured design specification — typography, color palette, layout grid, ingredient hierarchy, claim placement — that gets composed into a final design by a deterministic renderer. This separation lets us guarantee FDA compliance, enforce brand guidelines, and produce variations without re-rolling the dice.

For each prompt, the model outputs three to five variations across distinct stylistic axes (minimal vs. maximalist, clinical vs. lifestyle, etc.). Merchants pick a winner and our renderer produces the final print-ready files.

What we're working on next

The next frontier is brand-aware generation. We're training a fine-tuning layer that learns each merchant's aesthetic from their existing storefront — your product line, your colors, your tone — and applies it consistently across new SKUs. Early tests show this cuts the iteration count from 4-5 prompts down to one.

If you're curious to try the current model, install Productifi on your Shopify store. The first product is on us.