Ways to run the demo¶
GravyLensing can be set up in two physically different ways. They use the same software and the same settings, but they put the camera, the screen and the subject in different places, and they fail in different ways. Choosing between them is the first decision to make when planning a demo, because everything else — room layout, lighting, which lens mode to use — follows from it.
Mirror mode |
Telescope mode |
|
|---|---|---|
Camera points at |
The audience |
The screen |
Subject stands |
In front of the screen, facing it |
Between the camera and the screen |
Mirroring |
On, so movement matches |
Off |
Main weakness |
Bystanders in the background get lensed too |
Needs space, and careful field-of-view setup |
Best for |
Drop-in crowds, small spaces |
A controlled space, a faithful analogy |
Mirror mode¶
The camera sits next to or above the screen and looks back at the person, who watches themselves on the screen. The feed is mirrored horizontally so that moving left on the screen matches moving left in the room, exactly like a mirror. This is the intuitive setup: people work out what they are looking at in a second or two, and start playing with it immediately.
Turn mirroring on with Mirror camera feed in settings, or --flip on the
command line.
The catch is that everything the camera sees is a candidate mass. Anyone walking past behind the subject is detected as well, and the background image warps around them too. In a busy room the effect stops reading as “that person is bending spacetime” and starts reading as visual noise.
There are several options to push back on this, and they stack:
Focus mode (Track main subject only) keeps only the largest connected group of people in the mask and ignores bystanders elsewhere in the frame. Group distance controls how close two people must be, as a fraction of the frame width, to count as one group — raise it to keep a small group together, lower it to isolate one individual.
A region of interest (Shift+R) restricts detection to a rectangle of the frame. If the subject stands on a marked spot, an ROI around that spot removes the rest of the room from the problem entirely.
Colour mode sidesteps people altogether: hand the subject a brightly coloured object and lens on that. Bystanders are then irrelevant unless they happen to be wearing the same colour.
Person sensitivity can be lowered so that faint, distant or partially occluded figures fall below the detection threshold.
These help, but none of them beats physically isolating the subject. A backdrop behind the subject, a marked standing spot, a barrier, or simply pointing the camera slightly downward so that it sees floor rather than the queue, will do more for the demo than any setting. Treat the software options as a safety net for the people who wander in anyway, not as the plan.
Telescope mode¶
The camera points at the screen, and the subject stands in between, in the camera’s line of sight. What the camera sees is the screen — the distant source — with a person silhouetted against it. That silhouette becomes the lensing mass, and the screen shows the source deflected around them.
This is the faithful version of the analogy. It is the geometry of a real observation: a distant source, an intervening mass, and a telescope looking through the mass at the source. The audience is standing inside the light path, which is a much better story to tell than “the computer is drawing a warp around you”.
Mirroring should be off in this mode. The camera and the screen already face each other, so the image is not being used as a mirror and flipping it will make movement run the wrong way.
Telescope mode is far less prone to the background problem: the backdrop behind the subject is the screen itself, and anyone who is not between the camera and the screen is simply not in shot. In exchange it demands more from the space:
The camera must be far enough back. Too close and the screen does not fill the frame, so the illusion breaks at the edges where the camera sees the wall, the screen bezel, or the room beyond. Distance is what makes it feel seamless.
The field of view has to be matched to the screen. Aim for the screen to fill the frame as completely as you can manage, and use a region of interest (Shift+R) to trim whatever is left over around the edges.
The subject needs room to move without leaving the shot — and without standing so close to the camera that they fill the whole frame.
Lighting is a trade-off. A bright room lights the subject nicely but washes out the screen; a dark room does the reverse. The subject only has to be distinguishable from the screen behind them, not well lit.
Person detection is usually the safer lens mode here, because the backdrop is a moving, high-contrast image. Colour keying against a changing screen is much harder than colour keying against a wall.
Which one should you use?¶
If you have a small space, a drop-in audience, and no control over who walks through, use mirror mode with focus mode enabled and an ROI around the subject’s spot. It is more forgiving to set up and it explains itself.
If you have a dedicated space you can lay out in advance, use telescope mode. It is a better demonstration of the physics, and it solves the crowd problem by construction rather than by settings.
Setting either up in a real room is covered in Running a good demo.