From 4e7101c54ee1cb9f6493fd84172e23b881bf1a80 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?L=27=C3=A9lectron=20rare?=
 <108685187+electron-rare@users.noreply.github.com>
Date: Wed, 13 May 2026 23:58:11 +0200
Subject: [PATCH] fix(multi-hmr): NaN/Inf guard on v3d output

Multi-HMR PyTorch MPS sort occasionnellement des v3d avec NaN/Inf
ou vertices a magnitudes extremes (>5 m). Ces frames glissaient
sans guard jusqu'a AVLiveBody et apparaissaient comme pics/trous
sur le mesh.

Approche : np.isfinite check sur v3d + transl_np, skip person si
invalide. Plus sanity clamp sur |v3d|.max() > 5.0 m : humains ~2 m,
au-dela = garbage du modele, skip aussi. Le receiver Swift garde
le dernier good mesh pendant les frames skip via retain_window
500 ms (dc7de90).

Tradeoff : on perd la frame entiere du pid au lieu de juste les
verts NaN. Acceptable car corruption est globale au tensor v3d
quand elle arrive (jamais juste 2-3 verts isoles).
---
 data_only_viz/multi_hmr_worker.py | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)

diff --git a/data_only_viz/multi_hmr_worker.py b/data_only_viz/multi_hmr_worker.py
index 916d703..a00b6dc 100644
--- a/data_only_viz/multi_hmr_worker.py
+++ b/data_only_viz/multi_hmr_worker.py
@@ -369,6 +369,23 @@ class MultiHMRWorker:
                 shape_raw = hh["shape"].detach().cpu().numpy().flatten()
                 expr_raw = hh["expression"].detach().cpu().numpy().flatten()
 
+                # Skip persons with NaN/Inf vertices or transl : MPS can
+                # occasionally emit garbage that propagates to AVLiveBody
+                # as spikes / holes. We drop the frame for that pid and
+                # let the receiver's retain window keep the last good mesh.
+                if (not np.isfinite(v3d).all()
+                        or not np.isfinite(transl_np).all()):
+                    LOG.warning("Multi-HMR NaN/Inf at pid=%d, skipping", pid)
+                    continue
+                # Sanity clamp on extreme vertex magnitudes (humans are
+                # ~2 m ; vertices outside [-5, 5] m are model glitches).
+                if float(np.abs(v3d).max()) > 5.0:
+                    LOG.warning(
+                        "Multi-HMR v3d extreme |max|=%.1f at pid=%d, skipping",
+                        float(np.abs(v3d).max()), pid,
+                    )
+                    continue
+
                 pid_c = pid % self.num_persons
                 shape_n = min(10, len(shape_raw))
                 expr_n = min(10, len(expr_raw))