SAM 3D Body is a recently released human mesh recovery model that generates topologically coherent and visually clean meshes from a single image. The architecture relies on the Momentum Human Rig (MHR) for a low‑dimensional parametric representation, employs semantic‑invariant conditioning via DINOv3, and aligns training data through annotation‑based supervision. While it excels on typical poses and occlusions, systematic failures appear when reconstructing subjects with pronounced morphological variations such as geriatric muscle atrophy, scoliosis, or pregnancy; the output tends to smooth individual details toward an average shape. We attribute this to a perception‑distortion trade‑off: the low‑dimensional MHR limits fine‑grained biological detail, and the semantic‑invariant conditioning further enforces a regression‑to‑the‑mean effect. To overcome the limitation, we suggest two constructive directions: first, adopt implicit‑explicit hybrid representations to increase local geometric fidelity; second, incorporate Medical‑in‑the‑Loop alignment, using clinical measurements as additional constraints to push the model toward high‑precision medical use cases.\n\nReview