A convincing virtual set can fail at the edge of a sleeve. That is a good reason to begin a small virtual production setup with the person and the composite, rather than the most elaborate background the graphics card can render.
The UnityChromakey project offers a starting point for chroma keying and spill suppression inside Unity. NVIDIA's Broadcast tools suggest another direction for camera processing. These are useful leads for assembling a pipeline, although each approach has its own hardware and image-quality constraints.
The basic chroma-key problem is to estimate how much of each pixel belongs to the foreground. A clean green screen makes that estimation easier. Uneven lighting, shadows and reflective surfaces make it harder. Hair and motion blur produce mixed pixels, so the result should usually be an alpha value rather than a simple yes-or-no decision.
Removing green spill is a related task. The backdrop can illuminate a person's skin or clothing, leaving a green tint even where the foreground is fully opaque. Increasing the key threshold will not necessarily solve it. That can simply remove more of the person. Spill suppression needs to change the color while preserving the foreground's shape.
Then the virtual camera has to agree with the physical one. If the real image uses a wide field of view but the background is rendered as though seen through a long lens, the person seems pasted into a space with different rules. Camera height and horizon alignment are equally important, especially when the subject is supposed to stand on a virtual floor.
A locked camera makes the first version much simpler. The background can be framed once, and the composite can be inspected without also solving camera tracking. If camera movement is needed, the virtual viewpoint must follow it closely enough that the set does not slide behind the subject.
I would check timing with an intentionally obvious action, such as a hand moving across a visible virtual marker. The useful measurement is the complete path from capture to displayed composite. A fast keying shader cannot compensate for several frames waiting in a capture buffer.
For lighting, the aim is agreement rather than identical equipment. A bright virtual window on the left creates an expectation about the person's face. If the real face is lit strongly from the right, the viewer receives conflicting clues. A simpler background with lighting that can be matched may look more convincing than a sophisticated environment that cannot.
This is a pipeline I would build in layers: stable capture, usable matte, correct framing, then interaction and set detail. Each layer should have an image that makes its errors obvious. The reward for starting small is not merely a quicker setup. It is being able to identify why something looks wrong while there are still only a few possible causes.