Why decluttering a room turned out to be a harder AI problem than furnishing one, and why we shipped it as its own mode anyway.
I run ArchyBase — an AI room design tool — mostly solo, with a small team helping on the model side. I don't usually post here, but I'm trying to get better about writing down decisions while they're still fresh instead of only after they've been proven right or wrong. This one's about a feature that almost didn't ship.
For the first several months, ArchyBase only did one direction: you upload a photo of an empty or ugly room, and it fills it with furniture in whatever style you pick. That's the obvious version of "AI room design," and it's what everyone assumes the product does when they hear the pitch.
What I didn't expect was how many support messages were asking for the opposite. People uploading a photo of their current, fully furnished, lived-in room, and asking if there was a way to see it empty. Renters trying to picture the space without their stuff before deciding what to keep. People prepping a listing who didn't want to pay for a professional declutter shoot. A few just wanted to see their own room as a blank canvas before redesigning it, because starting from "here's my cluttered bedroom" makes it hard to evaluate a new layout honestly.
So we built the empty-room tool — decluttering as its own mode, not just a side effect of restyling.
Turns out it's the harder problem of the two, which is part of why I'd been avoiding it. Adding furniture to an empty room is, relatively speaking, forgiving — the model has a mostly blank surface to work with and a lot of room to be creative. Removing furniture means the model has to reconstruct what's actually behind the couch: the exact run of the flooring, the wall color and texture where a shelf was covering it, the baseboard hidden behind the bed frame. Get it slightly wrong and it's obvious immediately, because you're comparing against a real photo the person is looking at right now, not judging a fresh design on its own merits.
We shipped it with two explicit modes instead of trying to make one mode do both — declutter and stage — because early testing made it clear people wanted control over which direction they were going, not a single "make my room different" button. You can also nudge it with a short text prompt if the auto-detection guesses the room type wrong, and it'll give you a few variations instead of committing to one output, since "removing furniture" has more than one reasonable interpretation of what the floor probably looks like underneath a rug.
The thing I keep thinking about is that this wasn't a feature I planned. It came from noticing the same request showing up in support threads for a few weeks and finally asking "why do so many people want the tool to do less, not more." Most of the roadmap gets decided by looking at what the product should obviously do next. This one got decided by looking at what people kept asking for that wasn't on the roadmap at all.
Next thing I'm poking at: right now decluttering and restyling are two separate passes — empty it out, then (optionally) furnish it again in a new style. A few users have asked whether that should just be one step. I'm not sure yet — collapsing it might save a click, but it also removes the point where you get to actually look at the empty room, which several people told me was the part they cared about most. Might keep it two steps on purpose. Still deciding.
If you've used the empty-room mode and it did something weird with your floor or walls, I'd genuinely like to hear about it — that's the exact kind of feedback that gets the reconstruction model better.