EM Concepts — Dariusz Frączyk

Generated room scene in the refinement view, with a follow-up prompt refining the tile choice — A generated scene — refined with a follow-up prompt, one step from purchase.

The problem

Edward Martin’s room scenes are the heart of its product pages — photorealistic renders that let customers see tile, vanities, and decor in a real environment. Every one was hand-built by a small 3D team that couldn’t keep up with the catalogue. And even the best render shows a stylist’s room, not yours — the gap where hesitation and returns live.

The strategic bet: instead of scaling the 3D team, give customers a creative tool. Pick any scene on the site as a style reference, upload a photo of your own room, and generate that scene — same products, same style — in your space.

Before — product detail page with static room scenes

My role

Product Manager and UX Architect. I defined the product, planned the work, designed the UX architecture connecting the e-commerce experience with the generation experience, and directed my design team through execution.

The core design problem

Who is this for? A customer who has never touched an AI tool — and a design-savvy user who wants control. The interface had to work for someone who just wants to drag in a photo and see magic, without dumbing itself down for users who want to swap products or pull items from the catalogue. Serving one audience is easy. Serving both in one interface was the problem.

What didn’t work

We first explored a conversational UI — describe what you want, chat your way to a room. It demoed well and failed in practice.

Early conversational UI exploration

The solution: a prompt hero with progressive disclosure

One big, obvious starting point — with everything else layered behind it.

First visit: drag in a room photo, pick a reference scene, generate. Nothing else demands attention.

As you return: the interface reveals more. Swap a product in the generated scene. Add items from the Edward Martin catalogue. Refine style. The tool grows with the user instead of front-loading complexity.

Closing the loop: every generated scene tags its products. Like what you see in your own room — buy it from the image. Generation isn’t a toy bolted onto the store; it’s a path to purchase.

Project gallery with the prompt hero — a single input asking what would you like to create — First-run state: one obvious action. Space type, design style, featured products — nothing else demands attention.

Advanced controls popover — aspect ratio, resolution, and number of generations — Progressive disclosure, up close: plain-language prompt, reference chips, and the advanced controls that appear as the user grows into the tool.

Natural-language prompt describing the desired bathroom — Progressive disclosure, up close: plain-language prompt, reference chips, and the advanced controls that appear as the user grows into the tool.

Three concepts being generated simultaneously — Generation in progress — three concepts at once, so there is always a choice.

Key decisions

Prompt hero over conversation

A single visible action beats an open-ended chat. Customers shouldn’t have to know what to ask for.

Reference-based generation over free prompting

Starting from an existing scene anchors output quality and keeps results on-brand.

Progressive disclosure over feature parity upfront

Advanced controls exist, but earn their place on screen through use.

Outcome

The product was fully designed when the company hit financial headwinds and paused the launch. The design stands as a complete blueprint: a UX architecture that connects AI generation to commerce, scales content without scaling the 3D team, and stays usable for first-timers while rewarding power users.

Reflection

Designing for AI meant designing for uncertainty. The interface’s job isn’t to expose the model’s power — it’s to hide the model’s unpredictability behind choices the user already understands: a photo they own, a scene they like. Constrain the input, and you don’t have to apologize for the output.