Real-Time Physical Model

AI Workflow / 2024
Stable Diffusion · ComfyUI · Local Server
Role — Concept, Development & Visualization

Point a camera at a rough physical model—chipboard, foam, a few wood offcuts—and watch it become a finished, photoreal render in real time. Move a piece and the render moves with it. This was an experiment in collapsing the distance between a massing study you can touch and the image you'd normally wait hours to produce.

A live camera feed runs into a Stable Diffusion pipeline in ComfyUI, which reinterprets every frame against a prompt—material, context, light, lens—many times a second. The whole thing ran entirely on local office servers: the pipeline was served on the network at the workstation's IP address, and a short set of commands let internal staff connect—to drive it directly or to send their own live camera data in for processing.

It was built for big presentations. With the model on the table and the render on the room's main screen, clients could rearrange the massing with their own hands and generate their own renderings, live, in front of everyone.

01
Camera In A live DSLR or phone feed of a rough physical model streams straight into the pipeline.
02
Diffusion A ComfyUI / Stable Diffusion graph reinterprets each frame against a fixed prompt, in real time.
03
Instant Feedback Move, rotate, or swap a piece and the photoreal render updates with it—no export, no waiting.
04
On the Network Hosted on a local server and reachable by office users over its IP to operate it or feed live data.