April 24, 2026
From survey to six-pack?
A 3D Body from Eight Questions – No Photo, No GPU
No Photos, No GPU: 8 questions make your 3D body — commenters hype speed, roast the details
TLDR: A startup says it can build a 3D body from eight questions—no photos, no fancy hardware. Commenters love the speed and privacy, but argue about missing body details (like torso vs legs), nitpick the blog’s tone, and crack pocket-size jokes as e-commerce dreams meet real-world fit issues.
Forget photo shoots and gym selfies: a tiny neural network claims it can build your 3D body from just eight questions, in milliseconds on a regular computer. The makers brag about eyebrow‑raising accuracy and even model the “muscle weighs more than fat” problem. The crowd’s first reaction? Speed freaks shout “ship it!” with one user saying you could precompute millions of answers and serve results instantly. Privacy lovers cheer the no‑photo approach as a win for online shopping and virtual fitting rooms. A link‑dropper summed it up as paper + hustle = “we want to productize this.”
But the skeptics roll in. One critic slams a big blind spot: no question captures torso‑to‑leg ratios, a deal‑breaker for fit. Others poke at the blog’s vibe, wondering if the writing is AI‑assisted or just non‑native English—cue mini culture war. Meanwhile, fashion jokers hijack the thread: “Okay but… how big are the pockets?” which doubles as a jab at gendered sizing. The tension is classic: optimists see CPU‑only, question‑based sizing as an immediate e‑commerce superpower; purists argue without photos, some bodies will still be misread. It’s privacy vs precision, pragmatism vs perfection, with a side of meme. And yes, everyone wants those pockets measured, too.
Key Points
- •An eight-question questionnaire drives a small MLP that predicts 58 Anny body parameters without photos or GPUs.
- •The model reports ~0.3 cm height error, ~0.3 kg mass error, and ~3–4 cm MAE for bust/waist/hips, running in milliseconds on CPU.
- •The approach was inspired by height+weight regression (Bartol et al. 2022) but extends beyond h/w to capture shape differences.
- •Measurement convention differences (SMPL vertex landmarks vs ISO 8559-1 anatomy) significantly affected baseline error; bias correction reduced BWH MAE to ~7 cm on their data.
- •Training with a physics-aware loss helped fix a mass inconsistency in the Anny model and model the muscle-vs-fat effect.