Web configurator

Put a 3D object into a photograph of a real room

Status
Demo
Year
2026
The tool, end to end: background in, floor marked, model placed, render out.

Made with

Scene
glTF, loaded as .glb
Ground plane
Homography from four marked points
Render
ComfyUI, basic workflows for now
Test build
Vercel
Front end
Next.js and React, three.js in the viewport, Tailwind for the interface

What it does

A browser tool for putting a 3D object into a photograph of a real room. You drop in a background, mark where the floor is, add a .glb, and it renders the object into the scene at the right scale, in the right perspective, under the right light.

The point is to skip the shoot. If a product exists as a model, it can stand in a hundred rooms by tonight.

How it knows where the floor is

A photograph has no depth, so the tool has to be told where the ground is. You mark four points on the floor in the image, and they are matched to the corners of a flat rectangle. That pair gives a homography: a 3×3 matrix that maps any point on that plane in the photograph to a point on the plane in space, and back again.

Once it has that matrix the tool knows where the ground runs, how it recedes, and how large a meter is at any depth in the frame. The model then stands on the plane instead of floating in front of the picture.

Marking the floor. Until it is set, the background is flagged “needs floor”.

Where the render happens

Placement happens in the browser. The render happens on a ComfyUI backend. Those are basic workflows for now: enough to ground the object, match the light in the room and put a shadow where it touches the floor.

Before and after

Left is what the viewport gives you: the model lit flatly, floating, bounding box still showing. Right is what comes back from the render.

After. Before.
Office corridor. Two models, one render.
After. Before.
Same room, three models.
After. Before.
Different background, same placement pass.

Try it

A working build with a preset background. Rough edges included.

Open the test build

Next

  1. Automatic perspective detection

    Find the ground plane from the photograph, so nobody has to mark it.

  2. Multi-surface detection

    Tables, shelves and other planes, not just the floor.

  3. Object masking

    Let a model sit behind something that is already in the photograph.