I tried generating a Gaussian scene from a single ordinary photograph. My reaction afterward was that next time I should make a simpler face.
That joke is a fairly good description of what makes these experiments useful. A scene can look promising overall while one familiar detail makes the reconstruction feel wrong. Faces are particularly unforgiving because I know how the parts ought to relate, even when I cannot immediately describe the error.
A single photograph does not contain a complete 3D observation. It records the visible surfaces from one viewpoint. Any system that lets me move away from that viewpoint has to estimate geometry and make assumptions about what was hidden.
The result may be stored as Gaussians, but the storage format does not remove the missing information. 3D Gaussian splatting provides a way to represent and render a scene. Producing that representation from one image adds a reconstruction or generation problem before rendering begins.
I find it helpful to separate the input view from the new views when judging a result. Matching the photograph is one task. Remaining plausible when the camera moves is another. A small sideways movement can reveal stretched regions, inconsistent depth or surfaces that were never properly constrained by the input.
The face in my test was the detail that made me laugh, but the useful review should include less familiar areas too. Background objects can be wrong in ways that attract less immediate attention. A convincing face would not prove that the rest of the space was accurate.
For a creative effect, those approximations may be acceptable. A small amount of parallax can make an image feel more spatial without supporting a walk around the subject. The intended camera movement should match the evidence available to construct the scene.
I would rather keep an awkward result in the conversation than show only the angle that makes the method look perfect. It tells me where the representation stops being convincing and suggests a better next experiment: a clearer pose, simpler occlusion or a more limited viewing range.
The interesting part is being able to turn one photograph into something I can inspect from another position. The awkward part is useful too. It marks the boundary between the moment that was photographed and the space the system had to imagine.