Two identical 1.80 m people, one box-like picture, a room that is secretly skewed.
The science
The trick. A central collineation about the eye E maps every point P to P' = E + (P - E) s(P) with s = 1 / (1 + g (x + 0.03 (z - Ez))).
Planes go to planes and every ray through E is preserved, so the box room and the skewed room project to the same picture on the retina. Only the true distances change, by the factor s.
Size. A vertical object of height h at depth d (measured along the viewing axis) subtends an image height proportional to h / d. Two identical figures at depths dA and dB therefore have image-height ratio dB / dA. The visual system assumes a rectangular room, hence equal distances, and explains the size difference as a size difference of the people.
What the checks verify. (1) The projected corners of the skewed room coincide with those of the ordinary box. (2) Each of the six faces is exactly planar. (3) The measured image-height ratio of the two mannequins equals the depth ratio. (4) Each mannequin stands on the eye-ray of its ordinary-room position and really is 1.80 m tall.
Simplifications. Only geometry is modelled: one eye, no stereo, no head-motion parallax unless you move the head sliders (the illusion error grows as you do), no perceptual model of how strongly a real observer is fooled. The side-wall tilt coefficient 0.03 and the range of g are chosen for a pleasant room, not copied from a specific historical room. Idea credited to Adelbert Ames Jr.