The problem NeRFs are built to solve
Novel views from a handful of photographs
You have photographs from a few positions. You want a picture from a position you never stood in. Drag the white camera to a viewpoint between the captured ones and compare what naive blending produces against the truth — then turn on geometry and watch the difference.
Drop the count to 4 and sweep the camera around. Where you sit near a captured view the reconstruction is fine; between them it falls apart. That gap is exactly what a NeRF is trained to fill.
The error drops to zero exactly at each captured view — of course it does, that image is a photograph of that viewpoint — and rises in between. Add views and the peaks shrink. This scalloped curve is the sparse-view problem, and every technique in the field is a way of flattening it.
Why naive blending fails. Averaging two photographs pixel by pixel assumes that pixel i in one shows the same thing as pixel i in the other. It does not: move the camera and everything shifts across the frame, by an amount that depends on how far away it is. Near things slide a lot, distant things barely at all. Blending regardless produces two offset copies of the scene superimposed — the ghosting you can see in the strip above, and the same failure mode as the missing motion vectors in the DLSS demo.
Geometry is what fixes it. Switch the method to "using geometry" and the ghosting disappears. Now each pixel of the novel view is traced out to the point it actually sees, that point is projected back into each captured image, and the colour is taken from wherever it genuinely landed. Correspondence is established through the 3-D structure rather than assumed. The blend becomes correct because it is finally comparing like with like.
Which is the whole difficulty, because the geometry is not given. This demo cheats: it knows the scene, so it can look up exact depth. In reality you have photographs and nothing else. Photogrammetry attacks this by matching features and triangulating, and fails wherever matching fails. A NeRF attacks it from the other end — it never recovers geometry explicitly at all, but optimises a volume until its rendered images agree with the photographs. Geometry emerges as whatever makes the renderings correct.
And that reframes the sparse-view problem as a learning problem. With many views the constraints pin the field down almost completely. With few, many different fields explain the photographs equally well — the same ambiguity as super-resolution, one chapter section earlier, and answered the same way: with a prior about what scenes usually look like. Reduce the count to 3 or 4 and watch the error curve balloon between the captures. That gap is where all the research lives.