News for hackers. Stuff that matters.
Re:A 3D Body from Eight Questions – No Photo, No GPU by minhajulmahib on Saturday April 25, 2026 @10:09AM (#47901677)
The ancestry finding is the most honest part of this post — training on a uniform blendshape mix but inferring with the same fixed mix was essentially a 3 kg noise floor they built themselves. Elegant fix: just add ancestry to the questionnaire so train/inference distributions match. The physics-aware loss is interesting too. Including the Anny forward pass so mass gradients flow back through all volume-related params together — rather than solving each of the 58 outputs independently like Ridge — is exactly the right call. Ridge can’t couple params; the MLP hidden layers can. I’ve been thinking about a similar problem from the opposite direction — instead of reconstructing bodies, I’m working on running small models on very constrained hardware (NanoMind — 2GB RAM Android phones). The “boring model, interesting data pipeline” lesson resonates strongly. Upstream data quality always matters more than architectural complexity.
Re:A 3D Body from Eight Questions – No Photo, No GPU by gbgarbeb on Saturday April 25, 2026 @03:37PM (#47903934)
Grok, show me "Sexy Daniel."