Project Casper / ClockingMinds Research

Engines that remember the scene.

We are building state-driven engines that produce spatially aware images and video. Casper is the engine behind Reflection, and the reason a person, their home and a brand's product can share one believable frame.

Read the research See Reflection →

Update · Casper v0.9 · October 14, 2026

A first look at scene state

Today we're sharing an early research demo of Casper. From ordinary video shot on a single fixed camera, it builds a persistent model of the scene, and keeps track of what it saw, what it inferred, and what it does not know.

Casper v0.9 research demo, 0:31. Source footage on the left, the recovered scene on the right. Research preview, not a product.

What you're seeing

The input is a short clip of archival footage from one fixed camera. No depth sensor, no 3D capture, no extra views.
Casper turns it into a scene you can orbit, with the room, the counter and each person held as separate parts of the same space.
People are tracked as entities. When one moves out of view, Casper keeps a record of where they were instead of forgetting them, and flags any frame where what it expected and what it observed disagree.
Every surface carries its evidence: observed, retained but currently unseen, or inferred. Gaps stay gaps unless you choose to see the inferred fill, and that fill is always labelled.

Why it matters

Today's image and video models produce pixels. They don't hold a scene, so nothing stops a cup from changing hands or a wall from moving between frames. That is where many visual errors start.

Casper starts from the scene and treats each picture as a view of it. A system that knows what it has not seen can be told not to invent it.

For Reflection, this is the groundwork for keeping a product, a room or a person the same across every image and video made for one customer.

Limitations

Single fixed camera only. Moving cameras are not supported yet.
Geometry covers visible surfaces. What sits behind or beneath an object is unknown and shown as unknown.
Confidence shown in the demo reflects how well each surface is supported, not measured accuracy. We are not publishing accuracy figures yet.
This demo shows the scene state. Generating new images and video from that state is the next step and is not shown here.

What's next

Next we are extending Casper to moving cameras and multiple shots, and connecting the scene state to generation so outputs can be checked against it. We will share measured results when they are ready to be examined.

What makes it different

State first
The scene exists before the picture does.
Today's generators paint pixels and hope they agree. Casper keeps a persistent state of the scene: what is in it, where it sits and how it relates to everything else. Every frame is drawn from that state.
Spatially aware
Rooms keep their walls. Products keep their shape.
A chair stays the same size when the camera moves. A logo stays on the right pocket. Light comes from the window it came from in the last shot. Consistency is a property of the engine, not a lucky seed.
Images and video
One state, any number of views.
Because the scene is held, not reinvented, the same moment can be shown as a still, a new angle or a short clip without drifting between them.

Powering Reflection

Reflection decides what to show. Casper makes it hold together.

Your space: Products placed in a room that matches the one you shared.
Your look: The same person, recognisably, from shot to shot.
Their product: The brand's item exactly as it is in the catalog.

Casper is in active research. We share direction, not internals, and will publish measured results when they can be checked.