Back up geometry-warped face texture: docs/RECENT_EFFECTIVE_IMPROVEMENTS_20260822.md

experiments/h3d_original512_splat192_geometry_warped_face_texture_e260_20260824/docs/RECENT_EFFECTIVE_IMPROVEMENTS_20260822.md

@@ -1983,10 +1983,12 @@ remains the selected fallback.
 
 ### Neural Face Residual Renderer final epoch 265
 
-Status: **rejected**. The five-epoch branch completed normally, but the final
-result does not transfer Face-PSNR to SynBody and does not visibly recover
-identity-carrying high-frequency facial anatomy. Clean native-face e260 remains
-selected; e265 is retained only as a diagnostic ablation checkpoint.
+Status: **rejected as a standalone selection; retained as a composable backup**.
+The five-epoch branch completed normally, but the final result does not
+transfer Face-PSNR to SynBody and does not visibly recover identity-carrying
+high-frequency facial anatomy. Clean native-face e260 remains selected. The
+e265 full checkpoint and its 33-tensor residual-only adapter are preserved for
+future low-gate composition with a structural face-detail model.
 
 | Dataset / epoch | PSNR | FG-PSNR | LPIPS | alpha MSE | height ratio | native Face-PSNR | native face LPIPS |
 | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
@@ -2036,8 +2038,16 @@ neural_face_residual_e260_v1/train_1to1_e261_265_v1/validation/epoch_0265/synbod
 
 Checkpoint status: `checkpoint_epoch_0265/model.safetensors` is complete and
 readable (`1,737,631,564` bytes); both final fixed validations are complete and
-the source tmux exited cleanly. The fallback must resume clean non-residual
-e260, not this rejected checkpoint.
+the source tmux exited cleanly. New structural fallbacks must still resume
+clean non-residual e260. If a structural branch is later accepted, attach only
+the exported residual adapter with a low gate, freeze the structural path for
+calibration, and require an incremental dual-dataset gain before selecting the
+combined model.
+
+```text
+neural_face_residual_e260_v1/backup_adapter_e265/neural_face_residual_e265_adapter.safetensors
+neural_face_residual_e260_v1/backup_adapter_e265/neural_face_residual_e265_adapter.json
+```
 
 ### Geometry-Warped Neural Face Texture Renderer
 
@@ -2229,11 +2239,14 @@ error. Clean e260 remains selected pending the final e265 decision.
 
 ### Geometry-Warped Face Texture final epoch 265
 
-Status: **rejected**. The five-epoch experiment completed cleanly and produces
-small face-metric gains on both datasets, but identical fixed-sample visuals
-still show no meaningful recovery of missing facial anatomy or texture. The
-effect size is below a defensible paper claim, so clean e260 remains selected
-and the geometry-warped checkpoint is retained only as an ablation.
+Status: **accepted as the conservative metrics-first paper variant under the
+relaxed final threshold**. The five-epoch experiment completed cleanly and is
+the only structural branch with consistent native Face-PSNR and face-LPIPS
+gains on both datasets while keeping global changes at numerical-noise scale
+and introducing no visible regression. The effect remains subtle and must not
+be described as recovered facial anatomy. Clean e260 is retained as the
+no-risk production fallback; geometry-warped e265 is the selected novelty
+checkpoint for paper integration and further composition.
 
 | Dataset / epoch | PSNR | FG-PSNR | LPIPS | alpha MSE | height ratio | native Face-PSNR | native face LPIPS |
 | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
@@ -2268,7 +2281,10 @@ boundaries, nose and mouth anatomy, ears, hair edges, skin texture, and glasses
 stay blurred and far from GT. There is no seam, ringing, copied texture,
 view-inconsistent pattern, alpha drift, silhouette/scale change, or full-body,
 hands, or clothing regression. The mechanism transports information safely
-but the residual decoder converts it into only sub-visible corrections.
+but the residual decoder converts it into only sub-visible corrections. Under
+the final metrics-first criterion, that stability is acceptable: the claim is
+a safe source-to-surface refinement with small repeatable metric gains, not a
+dramatic visual-detail restoration.
 
 Inspected e265 artifacts:
 
@@ -2284,5 +2300,1111 @@ geometry_warped_face_texture_e260_v1/train_1to1_e261_265_v1/validation/epoch_026
 Checkpoint status: `checkpoint_epoch_0265/model.safetensors` is complete and
 readable (`1,738,429,028` bytes, 523 tensors); both final fixed validations
 are complete, the tmux exited normally, and all GPUs are free. The selected
-model remains clean non-residual e260. Next experiment: Wavelet-Constrained
-Multi-View Face Refiner from clean e260, never from this rejected branch.
+paper variant is geometry-warped e265; clean non-residual e260 remains the
+production fallback. Later failed branches remain useful controlled ablations.
+
+### Wavelet-Constrained Multi-View Face Refiner launch
+
+Status: **in progress; smoke accepted**. This fallback resumes the selected
+clean e260 generator and freezes all 438 inherited generator tensors. Four
+direct original-1024 source-face crops are encoded and camera-warped through
+the fixed 2,963-vertex SMPL-X face+ear surface, but the target decoder may
+predict only the three RGB Haar detail subbands of each `2x2` block. The Haar
+low-frequency coefficient, alpha, Gaussian geometry/opacity, and every pixel
+outside the feathered face ROI remain exactly equal to clean e260.
+
+The inverse-Haar residual has zero sum in every `2x2` block. A shared
+per-block headroom scale prevents RGB clipping without changing that zero-sum
+property. This removes the tone/contrast shortcut observed in the rejected
+Neural Residual and Geometry-Warped experiments while retaining learned
+source-view visibility/confidence attention, target-view conditioning, an
+uncertainty gate, native-GT high-frequency supervision, identity/delta
+regularization, and cross-view consistency.
+
+Accepted one-GPU forward smoke:
+
+- loaded `438/438` clean e260 tensors and rendered exactly `147,456` Gaussians;
+- initial refined/base maximum difference was `0` for both native faces and
+  full images;
+- Haar low-frequency absolute mean/max, alpha residual, and wavelet
+  coefficients were all exactly `0` at initialization;
+- source coverage mean/minimum was `0.98438 / 0.50000`, attention entropy/peak
+  was `0.98828 / 0.25794`, target coverage mean/fraction was
+  `0.17348 / 0.21649`, and every target vertex projected successfully;
+- finite forward peak memory was `8.78 GB`, with no projection, ROI, or
+  non-finite error.
+
+Accepted exact two-step backward smoke:
+
+- step-2 gradient norms were source encoder `2.98e-7`, visibility/confidence
+  fusion `1.77e-9`, decoder body `1.82e-6`, wavelet head `4.75e-4`, and gate
+  `9.98e-9`; frozen U-Net and Gaussian-head gradients were correctly absent;
+- the learned RGB high-frequency residual activated to mean/max
+  `3.70e-6 / 1.34e-4`, while the `2x2` low-frequency error remained only
+  `3.88e-14 / 3.64e-12` and alpha remained exactly unchanged;
+- source coverage stayed `0.97266`, all target vertices projected, gradients
+  and losses stayed finite, and peak memory was `8.88 GB`;
+- the step-2 fixed visual showed no ROI seam, ringing, floating contour,
+  alpha drift, or full-body change.
+
+Smoke artifacts:
+
+```text
+wavelet_face_refiner_e260_v1/forward_smoke/forward_smoke.json
+wavelet_face_refiner_e260_v1/forward_smoke/wavelet_runtime_audit.json
+wavelet_face_refiner_e260_v1/backward_smoke/gradient_audit.jsonl
+wavelet_face_refiner_e260_v1/backward_smoke/wavelet_runtime_audit.json
+wavelet_face_refiner_e260_v1/backward_smoke/images/step_0000002_face_pred_gt_error4x.png
+```
+
+Formal decision gate: train exactly e261-e265 on `7,200` MVHumanNet++ plus
+`7,200` SynBody samples per epoch and evaluate identical fixed samples at
+e261, e263, and e265. Accept only if both datasets gain face metrics and
+visibly sharper identity-carrying eye/nose/mouth/ear/hair/glasses detail with
+strictly unchanged low frequency, alpha, ROI exterior, scale, silhouette,
+hands, clothing, and full body. Otherwise retain clean e260 and record the
+wavelet mechanism as a paper ablation.
+
+### Wavelet-Constrained Face Refiner epoch 261
+
+Status: **in progress; currently below the acceptance bar**. The structural
+constraints are working exactly as intended, but the first epoch has not yet
+produced a transferable metric or visible-detail gain.
+
+| Dataset / epoch | PSNR | FG-PSNR | LPIPS | alpha MSE | height ratio | native Face-PSNR | native face LPIPS |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| MVHumanNet++ e260 clean | 27.96913 | 18.76916 | 0.015793 | 0.001984 | 1.00438 | 19.05802 | 0.263641 |
+| MVHumanNet++ e261 wavelet | 27.96910 | 18.76911 | 0.015793 | 0.001984 | 1.00438 | 19.05782 | 0.263397 |
+| SynBody e260 clean | 29.33703 | 20.99482 | 0.018124 | 0.002418 | 1.00020 | 20.55237 | 0.316577 |
+| SynBody e261 wavelet | 29.33706 | 20.99482 | 0.018124 | 0.002418 | 1.00020 | 20.55194 | 0.316455 |
+
+Against clean e260, MVHumanNet++ changes by `-0.00003` PSNR,
+`-0.00005` FG-PSNR, unchanged LPIPS/alpha/height, `-0.00020` native
+Face-PSNR, and `-0.00024` native face LPIPS. SynBody changes by
+`+0.00003` PSNR, effectively unchanged FG-PSNR/LPIPS/alpha/height,
+`-0.00043` native Face-PSNR, and `-0.00012` native face LPIPS. The frozen
+base metrics remain exactly clean e260.
+
+| Diagnostic | MVHumanNet++ e261 | SynBody e261 |
+| --- | ---: | ---: |
+| Source coverage mean / minimum | 0.99023 / 0.59375 | 0.99609 / 0.67500 |
+| Attention entropy / peak | 0.99121 / 0.25467 | 0.99648 / 0.25377 |
+| Target coverage mean / fraction | 0.16033 / 0.20212 | 0.18873 / 0.23741 |
+| Target projected fraction | 1.00000 | 1.00000 |
+| Uncertainty mean / max | 0.52344 / 0.95508 | 0.56367 / 0.97852 |
+| Haar coefficient absolute mean / max | 0.03683 / 0.88184 | 0.03967 / 0.94258 |
+| RGB residual absolute mean / max | 0.000094 / 0.01973 | 0.000133 / 0.02103 |
+| Haar low-frequency absolute mean / max | 9.35e-13 / 4.07e-10 | 1.37e-12 / 5.12e-10 |
+| Alpha residual max | 0 | 0 |
+| Gate / strength | 0.50056 / 1.0 | 0.50056 / 1.0 |
+| High-pass / identity / view loss | 0.008170 / 0.000094 / 5.87e-26 | 0.011009 / 0.000133 / 8.74e-26 |
+
+Identical fixed-sample visual verdict: refined and frozen-base columns remain
+nearly indistinguishable on frontal, profile, rear, occluded, and glasses
+views. No convincing eye, nose, mouth, ear, hair-edge, skin, or glasses detail
+has appeared relative to GT. Crucially, there is also no checkerboard, Haar
+ringing, ROI seam, copied texture, view inconsistency, alpha drift,
+silhouette/scale change, or full-body, hands, and clothing regression. The
+method is therefore mechanistically clean but visually inactive at e261.
+
+Inspected e261 artifacts:
+
+```text
+wavelet_face_refiner_e260_v1/train_1to1_e261_265_v1/validation/epoch_0261/mvhumannetpp/00_101010_face_pred_gt_error4x.png
+wavelet_face_refiner_e260_v1/train_1to1_e261_265_v1/validation/epoch_0261/mvhumannetpp/01_101020_face_pred_gt_error4x.png
+wavelet_face_refiner_e260_v1/train_1to1_e261_265_v1/validation/epoch_0261/synbody/00_090_face_pred_gt_error4x.png
+wavelet_face_refiner_e260_v1/train_1to1_e261_265_v1/validation/epoch_0261/synbody/02_092_face_pred_gt_error4x.png
+wavelet_face_refiner_e260_v1/train_1to1_e261_265_v1/validation/epoch_0261/mvhumannetpp/00_101010_input_pred_gt.png
+wavelet_face_refiner_e260_v1/train_1to1_e261_265_v1/validation/epoch_0261/synbody/00_090_input_pred_gt.png
+```
+
+Checkpoint status: `checkpoint_epoch_0261/model.safetensors` is complete and
+readable (`1,738,442,876` bytes), and both fixed-validation reports are
+complete. Training entered e262 on all eight GPUs at about `16.84 GB` per GPU
+with no OOM, non-finite, NCCL, projection, or ROI error. Clean e260 remains
+selected while the scheduled e263 evidence is pending.
+
+### Wavelet-Constrained Face Refiner epoch 263
+
+Status: **in progress; likely mechanism failure**. The detail coefficients and
+uncertainty gate are now strongly active, so the lack of improvement is not a
+zero-gradient or under-training artifact. Both datasets still lose a tiny
+amount of Face-PSNR, and the fixed samples show no identity-carrying detail
+recovery.
+
+| Dataset / epoch | PSNR | FG-PSNR | LPIPS | alpha MSE | height ratio | native Face-PSNR | native face LPIPS |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| MVHumanNet++ e260 clean | 27.96913 | 18.76916 | 0.015793 | 0.001984 | 1.00438 | 19.05802 | 0.263641 |
+| MVHumanNet++ e263 wavelet | 27.96907 | 18.76909 | 0.015800 | 0.001984 | 1.00438 | 19.05768 | 0.263092 |
+| SynBody e260 clean | 29.33703 | 20.99482 | 0.018124 | 0.002418 | 1.00020 | 20.55237 | 0.316577 |
+| SynBody e263 wavelet | 29.33702 | 20.99477 | 0.018124 | 0.002418 | 1.00020 | 20.55173 | 0.315967 |
+
+Against clean e260, MVHumanNet++ changes by `-0.00006` PSNR,
+`-0.00007` FG-PSNR, `+0.000008` LPIPS, unchanged alpha/height,
+`-0.00034` native Face-PSNR, and `-0.00055` native face LPIPS.
+SynBody changes by `-0.00001` PSNR, `-0.00005` FG-PSNR, unchanged
+LPIPS/alpha/height, `-0.00064` native Face-PSNR, and `-0.00061` native
+face LPIPS. The frozen base remains exactly clean e260.
+
+| Diagnostic | MVHumanNet++ e263 | SynBody e263 |
+| --- | ---: | ---: |
+| Source coverage mean / minimum | 0.99023 / 0.59375 | 0.99609 / 0.67500 |
+| Attention entropy / peak | 0.99121 / 0.25474 | 0.99648 / 0.25383 |
+| Target coverage mean / fraction | 0.16033 / 0.20212 | 0.18873 / 0.23741 |
+| Target projected fraction | 1.00000 | 1.00000 |
+| Uncertainty mean / max | 0.56055 / 1.00000 | 0.60469 / 1.00000 |
+| Haar coefficient absolute mean / max | 0.03003 / 0.94922 | 0.03217 / 0.97930 |
+| RGB residual absolute mean / max | 0.000123 / 0.02141 | 0.000160 / 0.02398 |
+| Haar low-frequency absolute mean / max | 1.21e-12 / 4.37e-10 | 1.65e-12 / 4.66e-10 |
+| Alpha residual max | 0 | 0 |
+| Gate / strength | 0.50172 / 1.0 | 0.50172 / 1.0 |
+| High-pass / identity / view loss | 0.008180 / 0.000123 / 8.74e-26 | 0.011017 / 0.000160 / 7.29e-26 |
+
+Identical fixed-sample visual verdict: refined and frozen-base faces are still
+nearly indistinguishable. There is no convincing recovery of eyes, nose,
+mouth, ears, hair edges, skin texture, or glasses. The coefficient maxima are
+near one and uncertainty maxima have reached one, yet the rendered residual
+is constrained to sub-visible, spatially balanced changes. No checkerboard,
+ringing, ROI seam, copied texture, alpha/low-frequency drift, silhouette or
+scale change, or full-body, hands, and clothing regression is visible.
+
+Inspected e263 artifacts:
+
+```text
+wavelet_face_refiner_e260_v1/train_1to1_e261_265_v1/validation/epoch_0263/mvhumannetpp/00_101010_face_pred_gt_error4x.png
+wavelet_face_refiner_e260_v1/train_1to1_e261_265_v1/validation/epoch_0263/mvhumannetpp/01_101020_face_pred_gt_error4x.png
+wavelet_face_refiner_e260_v1/train_1to1_e261_265_v1/validation/epoch_0263/synbody/00_090_face_pred_gt_error4x.png
+wavelet_face_refiner_e260_v1/train_1to1_e261_265_v1/validation/epoch_0263/synbody/02_092_face_pred_gt_error4x.png
+wavelet_face_refiner_e260_v1/train_1to1_e261_265_v1/validation/epoch_0263/mvhumannetpp/00_101010_input_pred_gt.png
+wavelet_face_refiner_e260_v1/train_1to1_e261_265_v1/validation/epoch_0263/synbody/00_090_input_pred_gt.png
+```
+
+Checkpoint status: `checkpoint_epoch_0263/model.safetensors` is complete and
+readable (`1,738,442,876` bytes), and both fixed validations are complete.
+Training entered e264 on all eight GPUs at about `16.84 GB` per GPU with no
+OOM, non-finite, NCCL, projection, or ROI error. Clean e260 remains selected;
+the run continues to the required final e265 for a complete ablation.
+
+### Wavelet-Constrained Face Refiner final epoch 265
+
+Status: **rejected; clean native-face e260 remains selected**. The required
+five epochs completed normally, but the final evidence confirms the e263
+mechanism diagnosis. Native face LPIPS improves by less than `0.00071` on
+both datasets while native Face-PSNR decreases on both, and identical fixed
+samples show no visible recovery of identity-carrying facial detail.
+
+| Dataset / epoch | PSNR | FG-PSNR | LPIPS | alpha MSE | height ratio | native Face-PSNR | native face LPIPS |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| MVHumanNet++ e260 clean | 27.96913 | 18.76916 | 0.015793 | 0.001984 | 1.00438 | 19.05802 | 0.263641 |
+| MVHumanNet++ e265 wavelet | 27.96905 | 18.76908 | 0.015800 | 0.001984 | 1.00438 | 19.05760 | 0.262970 |
+| SynBody e260 clean | 29.33703 | 20.99482 | 0.018124 | 0.002418 | 1.00020 | 20.55237 | 0.316577 |
+| SynBody e265 wavelet | 29.33701 | 20.99476 | 0.018124 | 0.002418 | 1.00020 | 20.55159 | 0.315869 |
+
+Against clean e260, MVHumanNet++ changes by `-0.00007` PSNR,
+`-0.00008` FG-PSNR, `+0.000008` LPIPS, unchanged alpha MSE and height,
+`-0.00042` native Face-PSNR, and `-0.00067` native face LPIPS. SynBody
+changes by `-0.00002` PSNR, `-0.00005` FG-PSNR, unchanged
+LPIPS/alpha/height, `-0.00079` native Face-PSNR, and `-0.00071` native face
+LPIPS. The frozen-base metrics remain exactly equal to clean e260.
+
+| Diagnostic | MVHumanNet++ e265 | SynBody e265 |
+| --- | ---: | ---: |
+| Source coverage mean / minimum | 0.99023 / 0.59375 | 0.99609 / 0.67500 |
+| Attention entropy / peak | 0.99121 / 0.25480 | 0.99648 / 0.25402 |
+| Target coverage mean / fraction | 0.16033 / 0.20212 | 0.18873 / 0.23741 |
+| Target projected fraction | 1.00000 | 1.00000 |
+| Uncertainty mean / max | 0.56348 / 1.00000 | 0.59102 / 1.00000 |
+| Haar coefficient absolute mean / max | 0.03027 / 0.97119 | 0.03258 / 0.99102 |
+| RGB residual absolute mean / max | 0.000132 / 0.023043 | 0.000158 / 0.026538 |
+| Haar low-frequency absolute mean / max | 1.30e-12 / 4.51e-10 | 1.61e-12 / 5.24e-10 |
+| Alpha residual mean / max | 0 / 0 | 0 / 0 |
+| Gate / strength | 0.50204 / 1.0 | 0.50204 / 1.0 |
+| High-pass / identity / view loss | 0.008181 / 0.000132 / 6.18e-26 | 0.011015 / 0.000158 / 6.77e-26 |
+| Wavelet low-frequency identity audit | 1.58e-9 | 3.45e-9 |
+
+Final visual verdict: the refined and clean-base columns are effectively
+indistinguishable across frontal, profile, rear, occluded, and glasses
+examples. Eyes, nose and mouth anatomy, ears, hair edges, skin texture, and
+glasses remain blurred relative to GT. There is no checkerboard, ringing,
+ROI seam, copied texture, view inconsistency, low-frequency or alpha drift,
+silhouette/scale change, or full-body, hands, and clothing regression. The
+method is stable and auditable, but it does not recover missing facial
+information.
+
+Inspected e265 artifacts:
+
+```text
+wavelet_face_refiner_e260_v1/train_1to1_e261_265_v1/validation/epoch_0265/mvhumannetpp/00_101010_face_pred_gt_error4x.png
+wavelet_face_refiner_e260_v1/train_1to1_e261_265_v1/validation/epoch_0265/mvhumannetpp/01_101020_face_pred_gt_error4x.png
+wavelet_face_refiner_e260_v1/train_1to1_e261_265_v1/validation/epoch_0265/synbody/00_090_face_pred_gt_error4x.png
+wavelet_face_refiner_e260_v1/train_1to1_e261_265_v1/validation/epoch_0265/synbody/02_092_face_pred_gt_error4x.png
+wavelet_face_refiner_e260_v1/train_1to1_e261_265_v1/validation/epoch_0265/mvhumannetpp/00_101010_input_pred_gt.png
+wavelet_face_refiner_e260_v1/train_1to1_e261_265_v1/validation/epoch_0265/synbody/00_090_input_pred_gt.png
+```
+
+Checkpoint status: `checkpoint_epoch_0265/model.safetensors` is complete and
+readable (`1,738,442,876` bytes). The e261 and e263 diagnostic checkpoints,
+all fixed validations, smoke audits, and implementation are preserved. The
+tmux exited normally, all eight GPUs are free, and there were no OOM, NaN,
+NCCL, projection, ROI, or checkpoint errors. The production selection remains
+the clean non-additive native-face e260 checkpoint.
+
+### Paper-ready face-detail ablation conclusion
+
+Status: **cascade complete; geometry-warped e265 selected as the modest
+metrics-first paper variant, with clean e260 retained as fallback**.
+The failures are mechanistically distinct and therefore useful as a controlled
+negative-results ablation rather than an undifferentiated hyperparameter
+search.
+
+| Candidate | Verified mechanism | Final diagnosis |
+| --- | --- | --- |
+| PatchGAN / DINOv2-GAN | Adversarial gradients and discriminator separation | Local sharpening did not transfer and global/face metrics regressed |
+| Residual Face-GS v1 | Additive 4,096-Gaussian face layer | Opacity collapsed, so combined and base-only renders became identical |
+| Face Attribute Refiner | In-place deltas on 4,096 existing face Gaussians | Active deltas did not produce transferable visible detail |
+| Native source-face parent split | 4-way face-Gaussian capacity and source features | Extra children did not recover semantic detail without introducing risk |
+| Surface Face-GS v2 | Independent SMPL-X-anchored surface layer | MVHumanNet++ alpha collapsed while SynBody retained harmful gray patches |
+| Neural Face Residual Renderer | Camera-conditioned bounded RGB/alpha ROI residual | Learned a tone/contrast shortcut: MVHumanNet++ improved but SynBody regressed |
+| Geometry-Warped Neural Texture | Explicit 1024-source feature transport through fixed face vertices | **Selected:** small dual-dataset face gains, numerically stable global output, no visible regression; no claim of anatomy recovery |
+| Wavelet-Constrained Refiner | High-frequency-only Haar correction with exact LL/alpha/exterior preservation | Tiny LPIPS gains, lower Face-PSNR, and no visible identity-detail recovery |
+
+The combined evidence rules out three easy explanations. Gradients were
+finite and nonzero, source coverage was high, and all target vertices
+projected; the branches were not simply disconnected. Capacity also became
+active: attention, uncertainty, residual, Gaussian, and wavelet statistics
+all moved substantially. Finally, the strict wavelet experiment proves that
+the earlier failures were not only caused by a tone/contrast shortcut.
+
+The remaining bottleneck is representational. Once the clean base render has
+lost eye boundaries, nose/mouth structure, ear shape, glasses edges, and skin
+texture, a sparse surface transport or bounded image-space residual cannot
+reconstruct those missing mid-frequency identity cues consistently across
+views and datasets. The wavelet branch further shows that zero-mean `2x2`
+detail corrections are too local to repair facial anatomy even when native
+1024 GT supervises them. The next defensible research direction is therefore
+an upstream, high-resolution canonical face representation, such as a dense
+SMPL-X UV-aligned identity feature field or face-specific high-resolution
+Gaussian plane decoded before rendering, with continuous surface
+rasterization and multi-view identity supervision. Further opacity tuning or
+post-render loss-weight searches are not supported by this ablation.
+
+Final selection: preserve and publish clean native-face e260 as the usable
+checkpoint. Preserve the rejected e265 checkpoints only for reproducibility
+and paper ablations; do not deploy or resume them.
+
+## Foveated face representation oracle cascade
+
+### Canonical UV face residual held-out oracle v0
+
+Status: **rejected**. A shared `512x512` SMPL-X UV residual atlas was optimized
+on three native target views and evaluated on the untouched fourth view for
+eight MVHumanNet++ and eight SynBody samples. The atlas has enough capacity to
+fit the supervised views, but the fixed SMPL-X correspondence does not
+transfer to the held-out camera.
+
+| Dataset | Seen Face-PSNR | Seen face LPIPS | Held-out Face-PSNR | Held-out face LPIPS |
+| --- | ---: | ---: | ---: | ---: |
+| MVHumanNet++ | +0.44800 dB | -0.07050 | -0.61885 dB | +0.04857 |
+| SynBody | +0.41565 dB | -0.06084 | -0.30409 dB | +0.02988 |
+
+The full-view average is misleadingly positive because three fitted views
+outnumber the one held-out view. The held-out regression on both datasets is
+the decisive result. This rejects a fixed canonical texture residual as the
+student target and isolates geometry/correspondence error as the remaining
+bottleneck. Alpha, height, and ROI-exterior pixels are unchanged, so the
+failure is not caused by silhouette or scale drift.
+
+Verified artifacts:
+
+```text
+workspace/h3d_original512_canonical_uv_face_oracle_v0/formal_8x2/metrics.json
+workspace/h3d_original512_canonical_uv_face_oracle_v0/formal_8x2/shard_*/mvhumannetpp/*_uv_base_gt_error4x.png
+workspace/h3d_original512_canonical_uv_face_oracle_v0/formal_8x2/shard_*/synbody/*_uv_base_gt_error4x.png
+```
+
+### Deformable Canonical Face Shell held-out oracle v0
+
+Status: **in progress; engineering smoke accepted**. The next oracle augments
+the same view-shared UV atlas with a bounded, view-shared deformation of the
+2,963 SMPL-X face-and-ear vertices. Displacements are limited to `12 mm` per
+axis and regularized by displacement magnitude, mesh adjacency smoothness,
+and edge-length distortion. The geometry is shared by all cameras, alpha and
+the clean e260 base remain fixed, and the fourth view is never optimized.
+
+The `2+2` sample, 40-step smoke produced finite nonzero offset gradients with
+mean displacement `1.25-1.42 mm`, maximum displacement `4.45-5.86 mm`, and
+peak memory `8.78 GB`. It did not saturate the bound or exhibit a disconnected
+geometry path. Its early held-out metrics remain negative, so the required
+`8+8`, 250-step formal oracle is running before a final mechanism verdict.
+
+```text
+workspace/h3d_original512_deformable_canonical_face_shell_oracle_v0/smoke_2x2/metrics.json
+workspace/h3d_original512_deformable_canonical_face_shell_oracle_v0/smoke_2x2/shard_*/mvhumannetpp/*_shell_base_gt_error4x.png
+workspace/h3d_original512_deformable_canonical_face_shell_oracle_v0/smoke_2x2/shard_*/synbody/*_shell_base_gt_error4x.png
+```
+
+Formal status: **rejected**. The full `8+8`, 250-step oracle confirms that
+target-side geometric freedom increases the seen/held-out gap rather than
+closing it.
+
+| Dataset | Seen Face-PSNR | Seen face LPIPS | Held-out Face-PSNR | Held-out face LPIPS |
+| --- | ---: | ---: | ---: | ---: |
+| MVHumanNet++ | +0.65049 dB | -0.07836 | -0.54202 dB | +0.05670 |
+| SynBody | +0.55897 dB | -0.06691 | -0.30564 dB | +0.03913 |
+
+Mean displacement is `2.55 mm` on MVHumanNet++ and `2.19 mm` on SynBody. The
+largest displacement-vector magnitudes are `18.08 mm` and `17.56 mm`, which
+remain within the configured per-axis `12 mm` bound. Alpha MSE and height are
+exactly unchanged. Visual inspection shows colored, anatomically misplaced
+detail on the optimized view while the held-out profile/rear views remain
+blurred or regress. The module is therefore active, finite, and geometrically
+bounded, but it still memorizes target-view evidence. A trainable student is
+not justified from this teacher.
+
+```text
+workspace/h3d_original512_deformable_canonical_face_shell_oracle_v0/formal_8x2/metrics.json
+workspace/h3d_original512_deformable_canonical_face_shell_oracle_v0/formal_8x2/shard_*/mvhumannetpp/*_shell_base_gt_error4x.png
+workspace/h3d_original512_deformable_canonical_face_shell_oracle_v0/formal_8x2/shard_*/synbody/*_shell_base_gt_error4x.png
+```
+
+Next gate: a Source-Observed Correspondence Field must source every detail
+value from the direct original-1024 input crops. It may optimize only bounded
+source-image sampling corrections and a view-shared surface detail gate on
+three target views, then must improve the untouched fourth target view on both
+datasets. This removes the target-atlas memorization path exposed above.
+
+### Source-Observed Correspondence Field held-out oracle v0
+
+Status: **rejected, but substantially more stable than target-side atlases**.
+The oracle samples only direct original-1024 source face crops, with bounded
+per-source/per-vertex correspondence offsets and one target-view-shared detail
+gate. It starts exactly at clean e260, uses three target views for optimization,
+and never exposes the fourth target GT to the parameters.
+
+| Dataset | Seen Face-PSNR | Seen face LPIPS | Held-out Face-PSNR | Held-out face LPIPS |
+| --- | ---: | ---: | ---: | ---: |
+| MVHumanNet++ | +0.06033 dB | -0.03645 | -0.01817 dB | +0.00903 |
+| SynBody | +0.10203 dB | -0.03778 | +0.02977 dB | +0.00517 |
+
+The average learned source correction is `1.66 px` for MVHumanNet++ and
+`1.51 px` for SynBody, the detail gate is active at `0.231/0.220` mean
+absolute magnitude, source coverage remains high, and gradients are finite.
+Alpha, height, and ROI exterior remain exact. Unlike the fixed/deformable UV
+teachers, held-out performance no longer collapses, which confirms that
+source-observed evidence removes the target-atlas memorization failure.
+However, held-out LPIPS still worsens on both datasets and fixed visuals do not
+show stable identity-detail recovery, so the mechanism is not yet suitable for
+student distillation.
+
+```text
+workspace/h3d_original512_source_observed_correspondence_oracle_v0/formal_8x2/metrics.json
+workspace/h3d_original512_source_observed_correspondence_oracle_v0/formal_8x2/shard_*/mvhumannetpp/*_correspondence_base_gt_error4x.png
+workspace/h3d_original512_source_observed_correspondence_oracle_v0/formal_8x2/shard_*/synbody/*_correspondence_base_gt_error4x.png
+```
+
+Mechanism diagnosis: a single corrected sample per SMPL-X surface point is too
+fragile when the parametric face and observed eyes, nose, lips, ears, hair, or
+glasses do not share exact subpixel correspondence. The next oracle uses
+visibility-aware soft attention over a bounded local epipolar neighborhood in
+each source crop, while retaining the same source-only evidence and held-out
+protocol.
+
+### Epipolar Foveated Feature Transport held-out oracle v0
+
+Status: **rejected**. The source-only transport was extended from one corrected
+sample to learned soft attention over a `3x3`, `+/-8 px` source neighborhood.
+The mechanism remained finite and active: mean offsets reached roughly
+`0.31-0.42 px`, maximum offsets approached `10.9 px`, attention peak remained
+about `0.87-0.88`, and peak memory was `8.79 GB`. It improved the three fitted
+views but did not transfer to the untouched fourth camera.
+
+| Dataset | Seen Face-PSNR | Seen face LPIPS | Held-out Face-PSNR | Held-out face LPIPS |
+| --- | ---: | ---: | ---: | ---: |
+| MVHumanNet++ | +0.04589 | -0.03158 | -0.01387 | +0.00661 |
+| SynBody | +0.09550 | -0.03471 | +0.04171 | +0.00070 |
+
+The global render remained stable and alpha/height were unchanged, but both
+datasets lost held-out perceptual quality. Local source attention therefore
+reduces correspondence error without solving semantic target-view ambiguity;
+it is not promoted to a trainable student.
+
+```text
+workspace/h3d_original512_epipolar_foveated_transport_oracle_v0/formal_8x2/metrics.json
+workspace/h3d_original512_epipolar_foveated_transport_oracle_v0/formal_8x2/shard_*/mvhumannetpp/*_correspondence_base_gt_error4x.png
+workspace/h3d_original512_epipolar_foveated_transport_oracle_v0/formal_8x2/shard_*/synbody/*_correspondence_base_gt_error4x.png
+```
+
+Next gate: Transmittance-Preserving Foveated Gaussian Subdivision replaces each
+selected face parent by four locally bounded children. Child optical-depth
+fractions sum exactly to the parent optical depth, and weighted position/color
+moments remain anchored, giving exact clean-e260 initialization without the
+opacity-collapse path of prior additive branches. A three-view-to-blind-view
+oracle is required before any native-256 U-Net fovea decoder is trained.
+
+### Transmittance-Preserving Foveated Gaussian Subdivision oracle v0
+
+Status: **rejected after the held-out smoke gate**. The mechanism itself is
+valid: 4,096 face parents are replaced by 16,384 children, giving 159,744
+Gaussians; child optical depths reproduce parent transmittance to within
+`3.8e-8`, weighted position/color moments remain anchored, initialization is
+clean-e260 equivalent, gradients are finite, and peak memory is `8.77 GB`.
+The additional capacity nevertheless fits the three exposed views without
+transferring perceptual detail to the blind camera.
+
+| Dataset | Seen Face-PSNR | Seen face LPIPS | Held-out Face-PSNR | Held-out face LPIPS |
+| --- | ---: | ---: | ---: | ---: |
+| MVHumanNet++ | +0.48774 | -0.04449 | +0.26626 | +0.01530 |
+| SynBody | +0.17051 | -0.02376 | -0.14148 | +0.01874 |
+
+```text
+workspace/h3d_original512_transmittance_foveated_subdivision_oracle_v0/smoke_2x2/metrics.json
+workspace/h3d_original512_transmittance_foveated_subdivision_oracle_v0/smoke_2x2/shard_*/{mvhumannetpp,synbody}/*_subdivision_base_gt_error4x.png
+```
+
+### Canonical Face Micro-TriPlane oracle v0
+
+Status: **rejected after the held-out smoke gate**. Two canonical face
+tri-plane scales (`32` and `64`) predict RGB detail at existing fixed Gaussian
+positions; xyz, scale, rotation, opacity, alpha, and ROI exterior remain fixed.
+The representation is finite and active but again learns a target-view texture
+field instead of a view-consistent identity signal.
+
+| Dataset | Seen Face-PSNR | Seen face LPIPS | Held-out Face-PSNR | Held-out face LPIPS |
+| --- | ---: | ---: | ---: | ---: |
+| MVHumanNet++ | +0.30105 | -0.04841 | -0.06631 | +0.03304 |
+| SynBody | +0.28370 | -0.03214 | -0.28463 | +0.03815 |
+
+```text
+workspace/h3d_original512_canonical_face_microtriplane_oracle_v0/smoke_2x2/metrics.json
+workspace/h3d_original512_canonical_face_microtriplane_oracle_v0/smoke_2x2/shard_*/{mvhumannetpp,synbody}/*_microtriplane_base_gt_error4x.png
+```
+
+### Native Feature-Indexed Foveated Gaussian Head oracle v0
+
+Status: **rejected, but selected as the most stable architectural precursor**.
+A zero-initialized convolutional head reads native `256` U-Net features and
+direct source-face high frequencies, then predicts only RGB corrections at the
+original `192^2` Gaussian indices. It preserves geometry/opacity and is much
+less prone to blind-view collapse than free per-instance fields, but it still
+misses the strict perceptual transfer gate.
+
+| Dataset | Seen Face-PSNR | Seen face LPIPS | Held-out Face-PSNR | Held-out face LPIPS |
+| --- | ---: | ---: | ---: | ---: |
+| MVHumanNet++ | +0.08185 | -0.01021 | +0.01731 | +0.00865 |
+| SynBody | +0.08150 | -0.00567 | -0.00635 | +0.00274 |
+
+Initialization is exactly clean e260, gradients through the feature head are
+finite, peak memory is `8.77 GB`, and global face deltas remain mildly positive
+(`+0.06794/-0.00550` MVHumanNet++; `+0.05608/-0.00357` SynBody). The remaining
+failure is specifically identity-carrying held-out detail. The next candidate
+therefore adds frozen semantic face priors rather than more target-side
+geometry, opacity, or texture capacity.
+
+```text
+workspace/h3d_original512_native_feature_foveated_head_oracle_v0/smoke_2x2/metrics.json
+workspace/h3d_original512_native_feature_foveated_head_oracle_v0/smoke_2x2/shard_*/{mvhumannetpp,synbody}/*_native_feature_base_gt_error4x.png
+```
+
+### Identity-Calibrated Semantic Foveation v0
+
+Status: **in progress; forward and exact two-step backward smokes accepted**.
+This branch resumes clean non-additive e260 and adds no discriminator or new
+render-time geometry. Frozen DINOv2 patch/class tokens supervise local semantic
+structure, while frozen ArcFace embeddings supervise identity only through a
+soft self-confidence gate. DINO patch errors are weighted by native-GT
+high-frequency saliency inside the direct original-1024 face mask. Only the
+original 438 U-Net plus Gaussian-head tensors are optimized.
+
+Forward smoke: 438/438 tensors loaded, exactly 147,456 Gaussians, finite
+semantic losses, exact off-axis projection counts, ArcFace cosine `0.83897`,
+identity confidence `0.98194`, and peak memory `8.95 GB`. The initial semantic
+weights were normalized before backward testing so the full-strength semantic
+term (`~0.0025-0.0030`) remains below the base training objective.
+
+Exact two-step backward smoke: U-Net input/mid/output gradient norms are finite
+and nonzero (`0.177/8.36e-5/0.325` at step 1 and
+`0.572/8.06e-5/0.229` at step 2); Gaussian-head gradients are `3.56/2.65`;
+peak memory is `26.9 GB`. No NaN, OOM, projection error, or trainable-count
+drift occurred. Formal e261-e265 is authorized with a 1,800-step warmup and
+strict cross-dataset acceptance.
+
+```text
+workspace/h3d_original_native512_splat192_identity_semantic_foveation_e260_v1/forward_smoke_v1/forward_smoke.json
+workspace/h3d_original_native512_splat192_identity_semantic_foveation_e260_v1/backward_smoke_v1/gradient_audit.jsonl
+workspace/h3d_original_native512_splat192_identity_semantic_foveation_e260_v1/train_1to1_e261_265_v1
+```
+
+#### e261 milestone
+
+Status: **in progress, but e261 does not pass the acceptance gate**. Training
+is healthy and has continued into e262. All eight GPUs remain active, peak
+memory is `34.1 GB`, the semantic warmup reached `1.0`, and no OOM, NaN,
+NCCL, projection, or checkpoint error was found. The semantic term remains
+active rather than collapsed (`~0.001-0.0046` in the end-of-epoch log), but
+the first formal validation does not yet transfer to both datasets.
+
+| Dataset | Metric | clean e260 | e261 | Delta |
+| --- | --- | ---: | ---: | ---: |
+| MVHumanNet++ | PSNR | 27.96913 | 27.93641 | -0.03271 |
+| MVHumanNet++ | FG-PSNR | 18.76916 | 18.76018 | -0.00898 |
+| MVHumanNet++ | LPIPS | 0.015793 | 0.015663 | -0.000130 |
+| MVHumanNet++ | alpha MSE | 0.001984 | 0.001986 | +0.000002 |
+| MVHumanNet++ | height ratio | 1.004377 | 1.003803 | -0.000574 |
+| MVHumanNet++ | native Face-PSNR | 19.05802 | 19.05807 | +0.00005 |
+| MVHumanNet++ | native face LPIPS | 0.263641 | 0.263794 | +0.000153 |
+| SynBody | PSNR | 29.33703 | 29.22126 | -0.11577 |
+| SynBody | FG-PSNR | 20.99482 | 20.99960 | +0.00478 |
+| SynBody | LPIPS | 0.018124 | 0.018201 | +0.000076 |
+| SynBody | alpha MSE | 0.002418 | 0.002452 | +0.000034 |
+| SynBody | height ratio | 1.000202 | 1.000756 | +0.000554 |
+| SynBody | native Face-PSNR | 20.55237 | 20.47973 | -0.07264 |
+| SynBody | native face LPIPS | 0.316577 | 0.317798 | +0.001221 |
+
+Identical fixed-sample visual inspection agrees with the metrics. MVHumanNet++
+sample `101020` is essentially unchanged, with only minute tone/detail shifts
+and no meaningful recovery of eyes, nose, mouth, ears, or hair. SynBody sample
+`098` is likewise visually unchanged, with small contour/tone differences but
+no identity-detail gain. No ringing, copied texture, or over-sharpening is
+visible. The e261 checkpoint is intact (`1,736,793,344` bytes) and retained as
+a diagnostic milestone; clean e260 remains selected.
+
+```text
+workspace/h3d_original_native512_splat192_identity_semantic_foveation_e260_v1/train_1to1_e261_265_v1/checkpoint_epoch_0261/model.safetensors
+workspace/h3d_original_native512_splat192_identity_semantic_foveation_e260_v1/train_1to1_e261_265_v1/validation/epoch_0261/mvhumannetpp/01_101020_face_pred_gt_error4x.png
+workspace/h3d_original_native512_splat192_identity_semantic_foveation_e260_v1/train_1to1_e261_265_v1/validation/epoch_0261/synbody/08_098_face_pred_gt_error4x.png
+```
+
+#### e263 milestone
+
+Status: **in progress, but e263 still fails the dual-dataset and visible-detail
+gate**. The full-strength semantic objective is active and numerically stable:
+over the logged e263 samples, semantic loss is `0.00103/0.00284/0.00507`
+(min/mean/max), strength is exactly `1.0`, and ArcFace cosine is
+`0.468/0.7719/0.973`. The mechanism has therefore not collapsed, but its
+benefit does not transfer consistently.
+
+| Dataset | Metric | clean e260 | e263 | Delta |
+| --- | --- | ---: | ---: | ---: |
+| MVHumanNet++ | PSNR | 27.96913 | 27.90536 | -0.06377 |
+| MVHumanNet++ | FG-PSNR | 18.76916 | 18.75566 | -0.01349 |
+| MVHumanNet++ | LPIPS | 0.015793 | 0.015762 | -0.000031 |
+| MVHumanNet++ | alpha MSE | 0.001984 | 0.001996 | +0.000012 |
+| MVHumanNet++ | height ratio | 1.004377 | 1.004289 | -0.000088 |
+| MVHumanNet++ | native Face-PSNR | 19.05802 | 19.10259 | +0.04457 |
+| MVHumanNet++ | native face LPIPS | 0.263641 | 0.264313 | +0.000671 |
+| SynBody | PSNR | 29.33703 | 29.29447 | -0.04256 |
+| SynBody | FG-PSNR | 20.99482 | 21.01558 | +0.02076 |
+| SynBody | LPIPS | 0.018124 | 0.018219 | +0.000095 |
+| SynBody | alpha MSE | 0.002418 | 0.002404 | -0.000014 |
+| SynBody | height ratio | 1.000202 | 1.000524 | +0.000322 |
+| SynBody | native Face-PSNR | 20.55237 | 20.48030 | -0.07208 |
+| SynBody | native face LPIPS | 0.316577 | 0.316724 | +0.000146 |
+
+On identical four-view face grids, MVHumanNet++ sample `101020` remains almost
+pixel-identical to e260: the `+0.04457 dB` Face-PSNR change does not recover
+eyes, nose, mouth, ears, or hair, and perceptual distance is slightly worse.
+SynBody sample `098` likewise shows no glasses, facial-feature, hair-edge, or
+profile recovery, matching its face-metric regression. Identical full-body
+grids show no obvious new scale, silhouette, hands, or clothing artifact, but
+also no useful visible gain. No ringing, copied texture, or over-sharpening is
+present.
+
+The e263 checkpoint is intact (`1,736,793,344` bytes), opens successfully with
+`471` stored safetensor keys after expected shared-LPIPS deduplication, and is
+retained only as a diagnostic milestone. Clean e260 remains selected. The
+formal branch continues through e265 before the final v0 verdict.
+
+```text
+workspace/h3d_original_native512_splat192_identity_semantic_foveation_e260_v1/train_1to1_e261_265_v1/checkpoint_epoch_0263/model.safetensors
+workspace/h3d_original_native512_splat192_identity_semantic_foveation_e260_v1/train_1to1_e261_265_v1/validation/epoch_0263/mvhumannetpp/01_101020_{input_pred_gt,face_pred_gt_error4x}.png
+workspace/h3d_original_native512_splat192_identity_semantic_foveation_e260_v1/train_1to1_e261_265_v1/validation/epoch_0263/synbody/08_098_{input_pred_gt,face_pred_gt_error4x}.png
+```
+
+#### e265 final
+
+Status: **rejected as a standalone improvement**. The five-epoch run completed
+cleanly through step `63000`. The semantic mechanism remained active at full
+strength: over logged e265 samples, semantic loss was
+`0.00108/0.00278/0.00511` (min/mean/max), strength was exactly `1.0`, and
+ArcFace cosine was `0.538/0.7899/0.989`. This is therefore a transfer failure,
+not an optimization collapse.
+
+| Dataset | Metric | clean e260 | e265 | Delta |
+| --- | --- | ---: | ---: | ---: |
+| MVHumanNet++ | PSNR | 27.96913 | 27.91741 | -0.05172 |
+| MVHumanNet++ | FG-PSNR | 18.76916 | 18.76046 | -0.00870 |
+| MVHumanNet++ | LPIPS | 0.015793 | 0.015686 | -0.000107 |
+| MVHumanNet++ | alpha MSE | 0.001984 | 0.002000 | +0.000017 |
+| MVHumanNet++ | height ratio | 1.004377 | 1.003895 | -0.000482 |
+| MVHumanNet++ | native Face-PSNR | 19.05802 | 19.09461 | +0.03659 |
+| MVHumanNet++ | native face LPIPS | 0.263641 | 0.263916 | +0.000275 |
+| SynBody | PSNR | 29.33703 | 29.28597 | -0.05106 |
+| SynBody | FG-PSNR | 20.99482 | 20.98476 | -0.01006 |
+| SynBody | LPIPS | 0.018124 | 0.018164 | +0.000040 |
+| SynBody | alpha MSE | 0.002418 | 0.002426 | +0.000008 |
+| SynBody | height ratio | 1.000202 | 1.000467 | +0.000265 |
+| SynBody | native Face-PSNR | 20.55237 | 20.49750 | -0.05487 |
+| SynBody | native face LPIPS | 0.316577 | 0.316724 | +0.000146 |
+
+Identical fixed-sample face inspection confirms the rejection. MVHumanNet++
+shows no meaningful recovery of eyes, nose, mouth, ears, or hair despite the
+small Face-PSNR increase, and face LPIPS is slightly worse. SynBody shows no
+recovery of glasses, facial structure, hair edge, or profile and regresses in
+both native face metrics. Full-body views remain visually stable with no new
+scale, silhouette, hand, or clothing artifact, but global PSNR decreases by
+about `0.051 dB` on both datasets. No ringing, copied texture, or
+over-sharpening was observed.
+
+The e265 checkpoint is intact (`1,736,793,344` bytes), opens with `471`
+safetensor keys, and is retained only for diagnosis. Clean e260 remains the
+selected checkpoint and the required initialization for subsequent structural
+correspondence branches; no later branch may resume from semantic e265.
+
+```text
+workspace/h3d_original_native512_splat192_identity_semantic_foveation_e260_v1/train_1to1_e261_265_v1/checkpoint_epoch_0265/model.safetensors
+workspace/h3d_original_native512_splat192_identity_semantic_foveation_e260_v1/train_1to1_e261_265_v1/validation/epoch_0265/mvhumannetpp/01_101020_{input_pred_gt,face_pred_gt_error4x}.png
+workspace/h3d_original_native512_splat192_identity_semantic_foveation_e260_v1/train_1to1_e261_265_v1/validation/epoch_0265/synbody/08_098_{input_pred_gt,face_pred_gt_error4x}.png
+```
+
+### Source-Anchored Multi-View Identity Cycle v1
+
+Status: **in progress; forward/backward smokes and strict held-out oracle
+accepted, formal e261-e265 launched**. This branch resumes clean e260 rather
+than any rejected checkpoint. It uses direct original-resolution source-face
+crops from the four diffusion views plus context view, frozen DINOv2 patch and
+class tokens, and frozen ArcFace embeddings. Visibility and pairwise agreement
+calibrate a source identity centroid; two random target views are constrained
+by saliency-weighted source-token matching and a cross-view identity cycle.
+Rear and occluded views are confidence-gated. The method retains all `147,456`
+Gaussians and trains the original `438` U-Net/Gaussian-head tensors without a
+trainable teacher, free xyz, or opacity changes.
+
+The one-GPU forward smoke loaded all `438/438` clean tensors, produced exactly
+`147,456 x 14` Gaussians, and had finite exact projections. Mean source/target
+valid fraction was `1.0`, source confidence was `0.82724`, DINO patch/class
+similarities were `0.51379/0.93543`, ArcFace similarity was `0.68562`, cycle
+loss was `0.003921`, and peak memory was `8.96 GB`. The exact two-step backward
+smoke produced finite nonzero gradients in U-Net input/mid/output and the
+Gaussian head on both steps; step-two maxima were
+`0.19585/7.37e-5/0.16290/1.59916`, with `27.19 GB` peak memory.
+
+The strict oracle optimized against three target views for eight steps while
+excluding the fourth view from every training loss. It is the first structural
+candidate in this cascade to improve both held-out Face-PSNR and held-out face
+LPIPS on both datasets.
+
+| Dataset | Subject | Base held-out Face-PSNR | Oracle | Delta | Base held-out LPIPS | Oracle | Delta |
+| --- | --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| MVHumanNet++ | 101010 | 16.37373 | 16.68758 | +0.31385 | 0.242798 | 0.241318 | -0.001480 |
+| SynBody | 090 | 18.65176 | 19.07337 | +0.42161 | 0.345304 | 0.341268 | -0.004036 |
+
+Identical held-out grids show subtle, spatially coherent changes around frontal
+eye/mouth contours, profile nose/lip boundaries, and the SynBody glasses/face
+edge. They do not show a broad tone/contrast shortcut, ringing, copied texture,
+or rear-view corruption. The visual gain is still modest after eight oracle
+steps, so this is a mechanism gate rather than formal acceptance.
+
+Formal training launched cleanly on eight GPUs with exactly `7,200`
+MVHumanNet++ and `7,200` SynBody groups per epoch, native `512` input, `192^2`
+Gaussian planes, `512` full render, two random `256` native face targets, and
+OneCycleLR max `5e-7`. At the first `100` e261 steps, cycle warmup progressed
+from `0.014` to `0.056`, source confidence remained `0.81-0.86`, logged peak
+training memory was `28.8 GB`, and no OOM, non-finite value, NCCL, projection,
+or ROI error was present. Clean e260 and the Neural Residual adapter backup
+remain preserved until e261/e263/e265 dual-dataset validation is complete.
+
+```text
+workspace/h3d_original_native512_splat192_source_anchored_identity_cycle_e260_v1/forward_smoke_v1
+workspace/h3d_original_native512_splat192_source_anchored_identity_cycle_e260_v1/backward_smoke_v1
+workspace/h3d_original_native512_splat192_source_anchored_identity_cycle_e260_v1/heldout_oracle_v1/{metrics.json,mvhumannetpp,synbody}
+workspace/h3d_original_native512_splat192_source_anchored_identity_cycle_e260_v1/train_1to1_e261_265_v1
+```
+
+#### e261 formal milestone
+
+Status: **e261 gate rejected; five-epoch branch remains in progress**. The
+checkpoint completed cleanly and is intact at `1,736,793,344` bytes with
+`471` readable safetensor keys. The cycle mechanism did not collapse:
+across 72 logged e261 points, source confidence was
+`0.578/0.8275/0.934` (min/mean/max), cycle loss was
+`0.00009/0.00257/0.00605`, and warmup strength reached `1.0`. Peak logged
+training memory remained `28.8 GB`, with no OOM, non-finite, NCCL,
+projection, or ROI error.
+
+| Dataset | Metric | clean e260 | e261 | Delta |
+| --- | --- | ---: | ---: | ---: |
+| MVHumanNet++ | PSNR | 27.96913 | 27.96440 | -0.00473 |
+| MVHumanNet++ | FG-PSNR | 18.76916 | 18.76505 | -0.00411 |
+| MVHumanNet++ | LPIPS | 0.015793 | 0.015770 | -0.000023 |
+| MVHumanNet++ | alpha MSE | 0.001984 | 0.001988 | +0.000004 |
+| MVHumanNet++ | height ratio | 1.004377 | 1.004673 | +0.000296 |
+| MVHumanNet++ | native Face-PSNR | 19.05802 | 19.13732 | +0.07929 |
+| MVHumanNet++ | native face LPIPS | 0.263641 | 0.265930 | +0.002289 |
+| SynBody | PSNR | 29.33703 | 29.29724 | -0.03979 |
+| SynBody | FG-PSNR | 20.99482 | 20.96430 | -0.03051 |
+| SynBody | LPIPS | 0.018124 | 0.018219 | +0.000095 |
+| SynBody | alpha MSE | 0.002418 | 0.002441 | +0.000023 |
+| SynBody | height ratio | 1.000202 | 1.000280 | +0.000077 |
+| SynBody | native Face-PSNR | 20.55237 | 20.42723 | -0.12514 |
+| SynBody | native face LPIPS | 0.316577 | 0.317188 | +0.000610 |
+
+Identical four-view grids show stable full-body scale, silhouette, hands, and
+clothing, with no ringing, copied texture, or rear-view corruption. However,
+MVHumanNet++ shows no meaningful eye/nose/mouth/ear recovery despite the small
+Face-PSNR gain, and its face LPIPS worsens. SynBody likewise shows no useful
+glasses, face, profile, or hair-edge recovery and regresses in both face
+metrics. The strict single-subject held-out oracle therefore has not yet
+transferred to the full fixed validation distribution. Training continues
+unchanged through e265 to distinguish an early-epoch fluctuation from a
+mechanism transfer failure; clean e260 remains selected.
+
+```text
+workspace/h3d_original_native512_splat192_source_anchored_identity_cycle_e260_v1/train_1to1_e261_265_v1/checkpoint_epoch_0261/model.safetensors
+workspace/h3d_original_native512_splat192_source_anchored_identity_cycle_e260_v1/train_1to1_e261_265_v1/validation/epoch_0261/mvhumannetpp/01_101020_{input_pred_gt,face_pred_gt_error4x}.png
+workspace/h3d_original_native512_splat192_source_anchored_identity_cycle_e260_v1/train_1to1_e261_265_v1/validation/epoch_0261/synbody/08_098_{input_pred_gt,face_pred_gt_error4x}.png
+```
+
+#### e263 formal milestone
+
+Status: **rejected at e263; branch continues to the required e265 final**.
+The checkpoint is intact at `1,736,793,344` bytes with `471` readable
+safetensor keys. The identity-cycle mechanism remains fully active rather than
+collapsed: over 72 logged e263 points, cycle loss was
+`0.00317/0.00494/0.00748` and source confidence was
+`0.620/0.8347/0.956` (min/mean/max), with strength fixed at `1.0`.
+Logged peak memory remains `28.8 GB`, and no OOM, non-finite, NCCL,
+projection, or ROI error occurred.
+
+| Dataset | Metric | clean e260 | e263 | Delta |
+| --- | --- | ---: | ---: | ---: |
+| MVHumanNet++ | PSNR | 27.96913 | 27.96043 | -0.00869 |
+| MVHumanNet++ | FG-PSNR | 18.76916 | 18.77522 | +0.00606 |
+| MVHumanNet++ | LPIPS | 0.015793 | 0.015701 | -0.000092 |
+| MVHumanNet++ | alpha MSE | 0.001984 | 0.001982 | -0.000002 |
+| MVHumanNet++ | height ratio | 1.004377 | 1.003901 | -0.000476 |
+| MVHumanNet++ | native Face-PSNR | 19.05802 | 19.10497 | +0.04695 |
+| MVHumanNet++ | native face LPIPS | 0.263641 | 0.265472 | +0.001831 |
+| SynBody | PSNR | 29.33703 | 29.36750 | +0.03047 |
+| SynBody | FG-PSNR | 20.99482 | 21.02378 | +0.02896 |
+| SynBody | LPIPS | 0.018124 | 0.018127 | +0.000003 |
+| SynBody | alpha MSE | 0.002418 | 0.002411 | -0.000007 |
+| SynBody | height ratio | 1.000202 | 0.999847 | -0.000355 |
+| SynBody | native Face-PSNR | 20.55237 | 20.47593 | -0.07644 |
+| SynBody | native face LPIPS | 0.316577 | 0.316138 | -0.000439 |
+
+The fixed grids again show no meaningful recovery of eyes, nose, mouth, ears,
+hair edge, skin, glasses, profile, or rear-view identity. MV improves
+Face-PSNR but worsens face LPIPS, while SynBody does the opposite. Full-body
+scale, silhouette, hands, and clothing remain stable and no ringing or texture
+copying is visible. The strict single-subject oracle therefore continues to
+fail distribution-level transfer. Clean e260 remains selected; e263 is
+diagnostic only.
+
+```text
+workspace/h3d_original_native512_splat192_source_anchored_identity_cycle_e260_v1/train_1to1_e261_265_v1/checkpoint_epoch_0263/model.safetensors
+workspace/h3d_original_native512_splat192_source_anchored_identity_cycle_e260_v1/train_1to1_e261_265_v1/validation/epoch_0263/mvhumannetpp/01_101020_{input_pred_gt,face_pred_gt_error4x}.png
+workspace/h3d_original_native512_splat192_source_anchored_identity_cycle_e260_v1/train_1to1_e261_265_v1/validation/epoch_0263/synbody/08_098_{input_pred_gt,face_pred_gt_error4x}.png
+```
+
+#### e265 final verdict
+
+Status: **rejected as a formal selection; diagnostic checkpoint preserved**.
+Training completed normally with `FORMAL_EXIT=0`. The final checkpoint is
+intact at `1,736,793,344` bytes with `471` readable safetensor keys. The
+mechanism did not collapse: over 72 logged e265 points, cycle loss was
+`0.00301/0.00497/0.00697` and source confidence was
+`0.666/0.8279/0.968` (min/mean/max), strength remained `1.0`, and peak
+memory was `28.8 GB`. No OOM, NaN, NCCL, projection, ROI, or checkpoint
+error occurred.
+
+| Dataset | Metric | clean e260 | e265 | Delta |
+| --- | --- | ---: | ---: | ---: |
+| MVHumanNet++ | PSNR | 27.96913 | 27.91753 | -0.05160 |
+| MVHumanNet++ | FG-PSNR | 18.76916 | 18.75223 | -0.01693 |
+| MVHumanNet++ | LPIPS | 0.015793 | 0.015808 | +0.000015 |
+| MVHumanNet++ | alpha MSE | 0.001984 | 0.002003 | +0.000020 |
+| MVHumanNet++ | height ratio | 1.004377 | 1.003901 | -0.000476 |
+| MVHumanNet++ | native Face-PSNR | 19.05802 | 19.15063 | +0.09261 |
+| MVHumanNet++ | native face LPIPS | 0.263641 | 0.265076 | +0.001434 |
+| SynBody | PSNR | 29.33703 | 29.29212 | -0.04492 |
+| SynBody | FG-PSNR | 20.99482 | 21.00132 | +0.00651 |
+| SynBody | LPIPS | 0.018124 | 0.018216 | +0.000092 |
+| SynBody | alpha MSE | 0.002418 | 0.002423 | +0.000005 |
+| SynBody | height ratio | 1.000202 | 0.999906 | -0.000296 |
+| SynBody | native Face-PSNR | 20.55237 | 20.52467 | -0.02771 |
+| SynBody | native face LPIPS | 0.316577 | 0.316431 | -0.000146 |
+
+Identical fixed grids show stable full-body scale, silhouette, hands, clothing,
+profiles, and rear views, without ringing or copied texture. They do not show
+meaningful recovery of eyes, nose, mouth, ears, hair edge, skin, or glasses.
+MVHumanNet++ trades a small Face-PSNR gain for worse face LPIPS and global
+metrics; SynBody trades a tiny face-LPIPS gain for worse Face-PSNR and global
+PSNR/LPIPS. The strict single-subject held-out oracle therefore did not
+transfer to the full validation distributions. Clean e260 remains selected;
+e265 is retained only for the paper ablation and mechanism diagnosis.
+
+```text
+workspace/h3d_original_native512_splat192_source_anchored_identity_cycle_e260_v1/train_1to1_e261_265_v1/checkpoint_epoch_0265/model.safetensors
+workspace/h3d_original_native512_splat192_source_anchored_identity_cycle_e260_v1/train_1to1_e261_265_v1/validation/epoch_0265/mvhumannetpp/00_101010_{input_pred_gt,face_pred_gt_error4x}.png
+workspace/h3d_original_native512_splat192_source_anchored_identity_cycle_e260_v1/train_1to1_e261_265_v1/validation/epoch_0265/synbody/00_090_{input_pred_gt,face_pred_gt_error4x}.png
+```
+
+### Zero-initialized Semantic Part-Fovea Adapter
+
+Status: **rejected by the held-out oracle; no formal fine-tune launched**.
+This parameter-efficient branch routes five zero-initialized rank-8 RGB
+experts for eyes, nose, mouth, ears, and hair edge using frozen DINOv2
+saliency. It changes only the RGB channels of existing Gaussians; geometry,
+opacity, alpha, and the clean full-body path remain exact.
+
+The forward smoke loaded all `438/438` clean e260 tensors, produced exactly
+`147,456` finite Gaussians, and had zero initial full-image, face, Gaussian,
+geometry, and opacity difference. Route coverage was `0.00816`, normalized
+route entropy was `0.94412`, and peak memory was `8.86 GB`. The exact
+two-step backward smoke produced finite nonzero gradients by step 2 in the
+adapter down projection (`2.66e-9`), spatial experts (`1.38e-9`), output head
+(`1.08e-5`), and audited clean Gaussian head (`6.43`), with `8.94 GB` peak
+memory.
+
+The strict 16-step three-view-to-held-out oracle nevertheless produced exact
+zero Face-PSNR and face-LPIPS change on both MVHumanNet++ subject `101010`
+and SynBody subject `090`. The adapter parameters moved, but its rendered RGB
+delta remained too small to survive BF16 Gaussian decoding/rasterization
+(`9.63e-8` mean on MVHumanNet++, `1.58e-7` on SynBody at step 16). This is an
+ineffective-capacity failure, not a numerical or projection failure. A zero
+metric change is explicitly not counted as a pass; clean e260 remains selected.
+
+```text
+workspace/h3d_original_native512_splat192_semantic_part_fovea_e260_v1/forward_smoke_v0
+workspace/h3d_original_native512_splat192_semantic_part_fovea_e260_v1/backward_smoke_v0
+workspace/h3d_original_native512_splat192_semantic_part_fovea_e260_v1/heldout_oracle_v0/metrics.json
+```
+
+### Visibility-Calibrated Canonical Face Token Transport
+
+Status: **rejected by the held-out oracle; no formal fine-tune launched**.
+This pre-render branch depth-filters direct original-resolution observations at
+the `2,963` SMPL-X face and ear vertices, fuses them into canonical surface
+tokens, maps the tokens to the `4,096` nearest existing face Gaussians, and
+predicts bounded RGB-only modulation. Geometry, opacity, alpha, and height are
+unchanged exactly. The zero-initialized modulator had finite gradients by step
+2 and used only `8.79 GB`.
+
+After 60 oracle steps, the mechanism was active (`0.01836`/`0.02448` mean RGB
+delta on MVHumanNet++/SynBody). Aggregate all-view metrics moved slightly in
+the desired direction on both samples, but strict held-out transfer failed:
+MVHumanNet++ held-out Face-PSNR changed `+0.08508 dB` while face LPIPS worsened
+`+0.006178`; SynBody changed `-0.08559 dB` and `+0.001072`. This confirms that
+visibility-filtered canonical color tokens still do not resolve the
+view-dependent correspondence ambiguity. The branch is retained as a paper
+ablation and is not promoted to formal training.
+
+```text
+workspace/h3d_original512_visibility_canonical_token_transport_oracle_v0/smoke_2x1/metrics.json
+workspace/h3d_original512_visibility_canonical_token_transport_oracle_v0/smoke_2x1/shard_00/mvhumannetpp/000_101010_canonical_base_gt_error4x.png
+workspace/h3d_original512_visibility_canonical_token_transport_oracle_v0/smoke_2x1/shard_00/synbody/000_090_canonical_base_gt_error4x.png
+```
+
+### Foveated Dual-Resolution Plane Adapter
+
+Status: **rejected by the held-out oracle; no formal fine-tune launched**.
+The prototype keeps the clean `192^2` Gaussian planes exact and samples direct
+native source-face RGB, mask, and high-frequency features as a sparse
+high-resolution face plane. A zero-initialized bounded adapter modulates only
+the RGB of `4,096` existing face-plane Gaussians. It had exact geometry,
+opacity, alpha, and height preservation, finite step-2 body/head gradients,
+active RGB deltas, and `8.79 GB` peak memory.
+
+The strict oracle rejected it. MVHumanNet++ held-out Face-PSNR/LPIPS changed
+`+0.00603 dB / +0.006294`, while SynBody changed
+`-0.01500 dB / -0.000159`. All-view SynBody face metrics improved, but the
+held-out split and MV perceptual metric did not. This is evidence that extra
+source-plane sampling resolution alone does not create a view-stable semantic
+face correspondence.
+
+```text
+workspace/h3d_original512_foveated_dual_resolution_plane_oracle_v0/smoke_2x1/metrics.json
+workspace/h3d_original512_foveated_dual_resolution_plane_oracle_v0/smoke_2x1/shard_00/mvhumannetpp/000_101010_canonical_base_gt_error4x.png
+workspace/h3d_original512_foveated_dual_resolution_plane_oracle_v0/smoke_2x1/shard_00/synbody/000_090_canonical_base_gt_error4x.png
+```
+
+### Visibility-Gated Face-Part Mixture of Experts
+
+Status: **rejected by the held-out oracle; no formal fine-tune launched**.
+Five zero-initialized low-rank experts specialize to eyes, nose, mouth, ears,
+and hair edge, with routing from projected face-part position, source support,
+and source-view identity. The mechanism was noncollapsed: routing entropy/peak
+were `0.239/0.180` on MVHumanNet++ and `0.349/0.261` on SynBody, RGB deltas
+were active, all geometry/opacity/alpha stayed exact, and peak memory was
+`8.79 GB`.
+
+All-view face metrics improved modestly on both samples, but the strict blind
+view did not: MVHumanNet++ held-out Face-PSNR/LPIPS changed
+`+0.01165 dB / +0.002245`; SynBody changed
+`-0.00361 dB / -0.000111`. Semantic specialization reduced the average
+perceptual error relative to the unpartitioned plane adapter, but did not make
+the transported color field consistently view-stable. The branch is retained
+as a noncollapsed negative ablation.
+
+```text
+workspace/h3d_original512_visibility_gated_face_part_moe_oracle_v0/smoke_2x1/metrics.json
+workspace/h3d_original512_visibility_gated_face_part_moe_oracle_v0/smoke_2x1/shard_00/mvhumannetpp/000_101010_canonical_base_gt_error4x.png
+workspace/h3d_original512_visibility_gated_face_part_moe_oracle_v0/smoke_2x1/shard_00/synbody/000_090_canonical_base_gt_error4x.png
+```
+
+### Source-Conditioned Frequency-Band Gaussian Modulator
+
+Status: **rejected by the held-out oracle; no formal fine-tune launched**.
+This RGB-only pre-render adapter predicts separately bounded mid/high-frequency
+residuals from native source-plane observations, gates high frequencies by
+observability, and removes each source plane's residual mean to preserve
+low-frequency color. The mechanism was active with exact geometry, opacity,
+alpha, and height preservation and `8.79 GB` peak memory.
+
+It improved all-view Face-PSNR/LPIPS by `+0.00636 dB / -0.000237` on
+MVHumanNet++ and `+0.04915 dB / -0.001873` on SynBody. However, both blind
+views worsened perceptually: held-out Face-PSNR/LPIPS changed
+`+0.00548 dB / +0.005833` and `+0.00538 dB / +0.001538`, respectively.
+Frequency isolation therefore prevents tone drift but does not by itself solve
+view-dependent face appearance. The next oracle explicitly models smooth
+target-view-conditioned reflectance rather than a single shared RGB value.
+
+```text
+workspace/h3d_original512_frequency_band_gaussian_modulator_oracle_v0/smoke_2x1/metrics.json
+workspace/h3d_original512_frequency_band_gaussian_modulator_oracle_v0/smoke_2x1/shard_00/mvhumannetpp/000_101010_canonical_base_gt_error4x.png
+workspace/h3d_original512_frequency_band_gaussian_modulator_oracle_v0/smoke_2x1/shard_00/synbody/000_090_canonical_base_gt_error4x.png
+```
+
+### View-Harmonic Face Reflectance Adapter
+
+Status: **rejected by the held-out oracle; no formal fine-tune launched**.
+This follow-up replaces a single shared face RGB correction with bounded
+first-order local tangent-frame reflectance coefficients. Existing face
+Gaussians receive a smooth target-camera-conditioned RGB residual while
+geometry, opacity, alpha, and height remain exact. The zero-initialized module
+had finite body/head gradients, active `0.01815`/`0.01748` mean RGB deltas,
+and `8.79 GB` peak memory.
+
+All-view face metrics improved slightly, but interpolation to the blind view
+failed on both datasets: MVHumanNet++ held-out Face-PSNR/LPIPS changed
+`-0.05398 dB / +0.009118`, and SynBody changed
+`-0.06024 dB / +0.006702`. A low-order smooth reflectance basis therefore
+does not recover the missing identity signal and overfits the three observed
+target directions. It is retained as a useful view-dependence ablation.
+
+```text
+workspace/h3d_original512_view_harmonic_face_reflectance_oracle_v0/smoke_2x1/metrics.json
+workspace/h3d_original512_view_harmonic_face_reflectance_oracle_v0/smoke_2x1/shard_00/mvhumannetpp/000_101010_canonical_base_gt_error4x.png
+workspace/h3d_original512_view_harmonic_face_reflectance_oracle_v0/smoke_2x1/shard_00/synbody/000_090_canonical_base_gt_error4x.png
+```
+
+### Final face-detail selection
+
+Status: **geometry-warped e265 selected under the final modest-gain
+criterion**. It provides the cleanest defensible novelty story: direct
+original-resolution multi-view face features are visibility/confidence fused
+on fixed SMPL-X face and ear vertices, deterministically warped into target
+face ROIs, and decoded as a bounded correction over the frozen clean e260
+generator. Both datasets improve native Face-PSNR and face LPIPS, while global
+changes are numerical-noise scale and fixed visuals show no regressions. The
+visual gain is sub-visible, so the paper must claim safe metric refinement and
+use the remaining branches as ablations rather than claim anatomy recovery.
+
+Selected checkpoint:
+
+```text
+workspace/h3d_original_native512_splat192_geometry_warped_face_texture_e260_v1/train_1to1_e261_265_v1/checkpoint_epoch_0265/model.safetensors
+```
+
+Production fallback:
+
+```text
+workspace/h3d_original_native512_splat192_native1024_face_e250_v1/train_1to1_e251_260_v1/checkpoint_epoch_0260/model.safetensors
+```
+
+### Foveated face representation oracle cascade
+
+Status: **in progress; capacity confirmed, view consistency not yet solved**.
+The clean native-face e260 checkpoint remains selected. These experiments are
+representation oracles and do not resume any rejected GAN or additive-opacity
+branch.
+
+#### Existing-Gaussian capacity oracle v0
+
+Status: **accepted as a capacity result only**. Per sample, the clean e260
+generator is frozen and the 4,096 closest existing face Gaussians are optimized
+against all four native 256 target views. Parent opacity is fixed; only xyz,
+scale, rotation, and RGB can change. The oracle passed its two-dataset face
+gate and used only `8.78 GB`, proving that the current Gaussian count can
+represent substantially sharper faces.
+
+| Dataset | Samples | Face-PSNR delta | Face LPIPS delta | PSNR delta | FG-PSNR delta | alpha MSE delta | height delta |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| MVHumanNet++ | 2 | +5.77504 | -0.16353 | +0.11217 | +0.44650 | -0.000079 | +0.02628 |
+| SynBody | 8 | +3.24450 | -0.10130 | +0.05718 | +0.26552 | -0.000035 | +0.01245 |
+
+Identical face grids show genuine recovery of eye boundaries, nose bridge,
+mouth shape, ears, and profile contour relative to e260, rather than only a
+tone shift. However, this all-view upper bound also changes silhouette/height
+on some samples and cannot establish novel-view consistency.
+
+Artifacts:
+
+```text
+h3d_original512_foveated_face_oracle_v0/formal_8x2/metrics.json
+h3d_original512_foveated_face_oracle_v0/formal_8x2/shard_00/mvhumannetpp/000_101010_optimized_base_gt_error4x.png
+h3d_original512_foveated_face_oracle_v0/formal_8x2/shard_01/mvhumannetpp/001_101020_optimized_base_gt_error4x.png
+h3d_original512_foveated_face_oracle_v0/formal_8x2/shard_06/synbody/006_096_optimized_base_gt_error4x.png
+```
+
+#### Three-view to held-out-view Gaussian oracle v1
+
+Status: **rejected; multi-view overfit diagnosed**. The formal gate uses eight
+samples from each dataset, rotates one of four target views out of optimization,
+and evaluates that view only after 250 steps. Seen views improve strongly, but
+the blind view regresses on both datasets.
+
+| Dataset | Seen Face-PSNR / LPIPS delta | Held-out Face-PSNR / LPIPS delta | Overall Face-PSNR / LPIPS delta |
+| --- | ---: | ---: | ---: |
+| MVHumanNet++ | +6.66871 / -0.16215 | -1.05600 / +0.10876 | +2.83100 / -0.09442 |
+| SynBody | +3.05064 / -0.09619 | -1.21999 / +0.09229 | +1.51131 / -0.04907 |
+
+The result rejects direct distillation of free per-instance Gaussian teachers:
+they fit target projections by moving/scaling/rotating face Gaussians but do not
+learn a transferable 3D face. The height-ratio drift (`+0.02064` MVHumanNet++,
+`+0.00818` SynBody) is consistent with that diagnosis.
+
+#### Fixed-geometry RGB-only held-out oracle v2
+
+Status: **rejected; correspondence failure remains without geometry motion**.
+All xyz, scale, rotation, opacity, alpha, and global outputs are fixed; only the
+same 4,096 parents' shared RGB is optimized on three views. Seen views still
+improve, but held-out views regress more strongly.
+
+| Dataset | Seen Face-PSNR / LPIPS delta | Held-out Face-PSNR / LPIPS delta |
+| --- | ---: | ---: |
+| MVHumanNet++ | +0.65346 / -0.07083 | -1.37106 / +0.07878 |
+| SynBody | +0.59687 / -0.04452 | -1.55330 / +0.05762 |
+
+This rules out free-geometry motion as the only cause. The existing Gaussian
+indices do not provide a sufficiently stable semantic surface correspondence
+for face-color transfer across views.
+
+#### Continuous SMPL-X surface pixel transport prototype
+
+Status: **rejected as direct RGB replacement; mechanism retained**. A new
+source-conditioned oracle rasterizes all 20,908 SMPL-X triangles for exact
+depth visibility, restricts transport to 5,742 face/ear triangles, and moves
+pixels from direct original-1024 source crops without optimizing target GT.
+The full-body depth buffer removes the initial through-head and through-clothing
+artifacts. One-sample-per-dataset smoke obtains Face-PSNR `+0.14082` on
+MVHumanNet++ and `+0.22461` on SynBody, but face LPIPS changes by `+0.03796`
+and `+0.01546`; direct replacement therefore remains misaligned at the
+identity-detail scale. A high-frequency-only variant also fails its smoke gate.
+
+The current next candidate is a shared bounded residual atlas in canonical
+SMPL-X UV space. It is optimized on three views and rendered into a fourth
+blind view, preserving e260 alpha, global geometry, ROI exterior, and low-risk
+base appearance. If fixed UV also fails, the evidence specifically motivates
+a learned deformable UV correspondence field rather than further Gaussian
+opacity or post-render loss searches.