Document which face path won, and the stochastic DPE eval

README.md

@@ -98,8 +98,25 @@ closeup, is held out as an unseen view, and the whole `C2` session is held out a
 session. The DPE codes were extracted from cam07 and cam30, both frontal tele views inside
 the training split, so no evaluation view leaks into the face conditioning.
 
+## Which face path won
+
+On P1, with identical data, schedule and architecture, `P1_smplx` **beat** `P1_dpe`. That is
+why the other released avatars use the SMPL-X path. `P1_dpe` is the paper-faithful
+formulation, and the right choice when you have no trustworthy face fit, but it is not the
+stronger one on this data.
+
+Read that carefully, though: whole-image and even head-crop metrics could not separate the
+two at all, differing only in the fourth decimal. A face ablation moves roughly 1% of the
+pixels, so whole-image PSNR is dominated by torso and clothing. Only a mouth region defined
+from the jaw-driven SMPL-X vertices distinguished them (PSNR +0.58, SSIM +0.015, LPIPS
+-10%). It is one subject, so treat it as a direction, not a settled result.
+
 ## Limitations
 
+- **`P1_dpe` evaluation is not deterministic.** The per-frame face code is a random convex
+  combination of the two camera codes, redrawn on every sample, and that path runs at eval
+  time as well as during training. Restrict the evaluation to one camera per frame if you
+  need reproducible numbers.
 - One subject per avatar. These are personalised avatars, not a generalisable model.
 - Driving signal must be a registered SMPL-X sequence in the DREAMS-AVATAR convention.
 - Densification is disabled during training, and there is no oral-cavity geometry, so teeth