I've been heads-down for the last few months on Deroom AI, an AI room design tool that turns a single photo of your real room into a photorealistic redesign in 30 seconds. The 0→1 build was way harder than I expected and along the way I changed my mind about a bunch of things. These are the four that cost me the most time.
1. Generic image generation is the wrong primitive
The first version was a thin wrapper around a stable-diffusion checkpoint with a prompt template. It produced beautiful kitchens. They just weren't your kitchen. Walls moved. Windows shifted. Cabinets rearranged themselves.
That's useless if the entire reason a homeowner is on the site is to decide whether to repaint their existing cabinets sage green or off-white.
The fix took weeks: switch to a controlnet-driven pipeline that uses the input photo as a structural reference. Footprint, plumbing, doors, windows, ceiling height — locked. Only finish changes (paint, tile, cabinetry, lighting, soft furnishings).
Lesson: if the user can't recognize the output as a transformation of their input, the product is a different product.
2. One knob ("redesign") covers maybe 30% of intent
I shipped a single "redesign this room" mode. Logs showed people generating 3-5 variations and bouncing — they couldn't get what they actually wanted, which was usually a smaller intervention.
Now there are five action modes:
- Full Redesign — everything changes. The original use case.
- Quick Refresh — keep the bones, swap finish. The $1K weekend project.
- Declutter — strip everything except structure. Great for real-estate listings.
- Recolor Only — only paint and material colors change. Test 10 paint palettes in 5 minutes.
- Furniture Swap — same room, different sofa or bedframe. Useful before ordering.
The Recolor and Furniture Swap modes were ~3 weeks of additional pipeline work each (different model conditioning, different prompts) but they unlocked the largest category of users — people who don't actually want a "redesign", they want help with one specific decision.
3. Style transfer needs per-room calibration
I assumed "Japandi" was Japandi. It isn't. Japandi in a kitchen looks completely different from Japandi in a bedroom — different cabinetry, different surfaces, different lighting expectations.
The first generation of style prompts produced uncanny output: a "Mediterranean" bedroom that looked like a hotel lobby, a "Modern" closet that looked like an empty warehouse.
I ended up calibrating each of 12 styles separately for each room type (kitchen, bathroom, bedroom, living, office, closet, exterior, garden, landscape, etc.) instead of having one global style prompt. The matrix is way bigger than I planned for. But the output quality jumped a tier.
Side effect: it makes per-room landing pages much more useful. The ai bedroom design page can show actual bedroom variations, not generic interior renders, because the system was built to produce bedroom-specific output.
4. Pricing — credit-based, not subscription-based, was the right call
Original plan: tiered subscriptions with monthly generation caps (50/200/600).
Problem: people don't think in "generations per month". They think in "I have a kitchen project this month, I need ~30 generations this week, then nothing for two months." Monthly resets felt punitive.
Switched to a credit model: every generation = 10 credits, plans give you a credit allowance, no use-it-or-lose-it for first 30 days, predictable cost per output. Conversion went up.
This is unrelated to the AI part of the product but it was probably the single biggest revenue lever I changed.
What I'd build next
The thing I keep wanting to ship and haven't is side-by-side variations. Right now you generate one image at a time. The natural mental model is "show me 4 versions of the same room in 4 different paint colors", and we don't quite do that yet — you have to run 4 generations and stitch them in your head.
Architecturally it's not hard, but the UX is finicky. Probably next sprint.
If any of this resonates and you want to play with the workflow, deroomai.com is free to try (10 credits on signup, no credit card). Would love to hear what you build with it — or what I'm getting wrong.
Top comments (0)