From Photographs to Patterns: Production GPU Pipelines for Monocular Body Measurement and Diffusion Based Virtual Try On

Garment retail and bespoke tailoring depend on accurate body measurements, yet the tape measure remains the only instrument most customers can reach. This preprint reports two GPU services built to replace it for a useful class of cases: an anthropometric service that estimates twelve body measurements from a single casual photograph and a stated stature, and a virtual try on service that renders a chosen garment onto the same photograph. The measurement service began as a two dimensional pipeline that approximated girths as ellipses. It produced errors of roughly four to five inches, for structural reasons: human cross sections are not elliptical, two dimensional landmarks mark joint centres rather than body surfaces, and a single view cannot resolve depth. It was rebuilt on monocular three dimensional mesh recovery (SAM 3D Body, Momentum Human Rig), with girths taken as polygonal perimeters of mesh cross sections. An ablation on the exported reference mesh quantifies what each design decision is worth. Canonical body frame alignment dominates: at the 13.2 degree lean measured for the reference subject, leaving it uncorrected inflates chest, waist and hip girths by 3.1, 3.6 and 2.0 per cent. Arm removal is a precondition, since only 68 per cent of torso slice heights yield a single closed loop. Chest girth moves 0.98 cm per centimetre of placement error, against 0.20 cm for the waist. The convex hull envelope changes no girth by more than 0.22 cm and is a safeguard rather than a major correction. The try on service uses CatVTON, selected after a review of six diffusion based alternatives under an eight gigabyte memory ceiling. Both services are packaged as containerised serverless GPU workers. Steady state latency is ten to fourteen seconds for measurement and seventy one to seventy three seconds for try on. Preprint. Not peer reviewed. Validation is single subject and preliminary: one neck measurement agrees with a tape reference to within 0.4 cm, and population scale validation remains outstanding. See Sections 6.3 and 9 of the paper. Worker implementations are not publicly released.