What volumetric means
A flat image with a depth map gives you a little parallax and falls apart at the edges of objects, because there is nothing behind the thing in front. Move far enough and you see the seam — a stretched smear where the occluded surface should be.
Volumetric content has actual extent. Move and the occlusion changes correctly, because something genuinely exists behind the foreground. That is the difference, and it is not a matter of degree.
It is expensive in every dimension — capture, storage, bandwidth, compute and authoring time — which is why most media marketed as "3D" or "immersive" is the first thing rather than the second, and why the gap between the demo and the shipped experience is usually so large in this category.
| Depth map on a photo | Volumetric | |
|---|---|---|
| Small head movement | Convincing | Convincing |
| Moving around an object | Breaks at the silhouette | Holds |
| Occlusion | Faked, fails at edges | Correct |
| File size | Small | Large, often by orders of magnitude |
| Authoring cost | Low | High |
The approach here
Consistency first. Before depth is worth anything, the same scene has to be reproducible from several angles — same objects, same light, same materials. That is the problem the graph is built for and the part that works today.
Depth without that consistency produces a scene that reorganises itself when you look away, which is worse than a flat image rather than better. The viewer forgives a flat image for being flat. They do not forgive a space for being unreliable, because it breaks the one promise the medium makes.
So the order of work is consistency, then depth, then format. Anyone doing it in the other order is building a demo.
| Status | |
|---|---|
| Multi-view scene consistency | Works today |
| Shared lighting across views | Works today |
| References that hold a subject across shots | Works today |
| Consistent depth between views | Partial — visual agreement, no computed geometry |
| Volumetric capture ingest | Not in this build |
| Point cloud or gaussian output | Not in this build |
| Headset-native export | Not in this build |
Questions
Can I bring in a gaussian splat or a photogrammetry scan?
Not as a scene format. You can use rendered views from one as reference images, which is a normal way to hold a real place.
Is generative volumetric capture realistic soon?
Generating consistent novel views is advancing quickly. Generating a volumetric representation that holds up under free movement is a harder problem, and anyone giving you a confident date is guessing.
What should I do today if volumetric is my deliverable?
Capture volumetrically with the tools built for it, and use this for concept, look development and previs. That division of labour is honest and it is what we would advise even if it were not in our interest.