How Space Remodel works

TL;DR

Space Remodel uses a diffusion model conditioned on the architectural geometry of your input photo (via ControlNet + M-LSD edge detection) to preserve the bones of your room while reimagining the furniture, materials, and lighting. The result is photoreal, structurally faithful, and ready in 8–20 seconds.

The problem with naive image generation

If you hand a generic diffusion model a photo of your living room and ask for “Scandinavian style”, you’ll get a Scandinavian living room — but not your living room. Walls move, windows shift, the layout reshuffles. That’s useless for anyone trying to visualize a redesign.

ControlNet: keeping the bones

ControlNet is a neural network architecture that lets you steer a pretrained diffusion model with additional spatial inputs — edge maps, depth maps, segmentation masks. Space Remodel uses ControlNet conditioned on M-LSD-detected straight lines from your photo. The model treats those lines as load-bearing — walls, floors, windows, doorways stay put. Everything else is free to change.

Training data: 50k+ interior photos from Places2 + descriptions from BLIP image captioning, trained on 8× A100 80G GPUs for 150 GPU-hours. The Canny edge model is used as a starting checkpoint.

What you get to control

• Style — 45+ design styles (Scandinavian, Industrial, Japandi, Mid-century Modern, etc.).
• Room type — 22+ categories (Living Room, Kitchen, Bedroom, Wine Cellar, etc.). Influences furniture priors.
• Color palette — optional override; otherwise picked by the style.
• Materials — flooring, wall finishes, key furniture surfaces.
• Render strength — preserve more of original (light retouch) vs reimagine from scratch.

What you don’t control (by design)

• Wall, window, and floor positions — preserved by ControlNet.
• Lighting direction — matched to your photo’s natural lighting.
• Architectural style — Space Remodel reimagines decor, not the building.

4K output

Pro renders are upscaled to 3840×2160 using a second-stage diffusion pass tuned for interior photography. Sharpness is good enough for real-estate listing photos and large-format print.

On-device acceleration

On Apple silicon Macs and Vision Pro, Space Remodel performs the initial preview stage on-device using Core ML. The final 4K render runs server-side for quality reasons. The hybrid pipeline keeps perceived latency low.

Privacy

Uploaded photos are processed in transit and discarded after the render session. They are not used to train future models. See Privacy.

Try it

Download Space Remodel free for iPhone, iPad, Mac, and Vision Pro.