fix VAE latent dims: 32 tokens x 32-d (1,024-d total), not 32x1024

README.md

@@ -24,7 +24,7 @@ Single-step, scaffold-dominant diffusion for 3D scene completion on SemanticKITT
 
 ### SMC 2026 (accepted) — *CloudDiffusion: Diffusion-Based Scene Completion in the Point Cloud Domain*
 The companion short paper covers:
-- **Multi-token Gaussian VAE** (7.1 M params): residual PointNet encoder + 32-query cross-attention pooler producing 32 × 1024-d Gaussian latent tokens; transformer decoder with 5 cross-attention blocks reconstructs 8,000 scene points. Squared CD **0.120 ± 0.026 m²** at **1.6 ms / frame**, no codebook collapse (vs ~16 m² for the VQ-VAE alternative).
+- **Multi-token Gaussian VAE** (7.1 M params): residual PointNet encoder + 32-query cross-attention pooler producing 32 × 32-d Gaussian latent tokens (1,024-d latent total); transformer decoder with 5 cross-attention blocks reconstructs 8,000 scene points. Squared CD **0.120 ± 0.026 m²** at **1.6 ms / frame**, no codebook collapse (vs ~16 m² for the VQ-VAE alternative).
 - **Anchor-based ICP ground-truth refinement (v2 GT)**: per-scan ICP against a w=17 temporal window (±17, 35-scan context) with displacement-gated acceptance (τ = 0.15 m), SOR/ROR cleanup, 0.1 m voxelisation, over 23,201 frames (considerably denser than the original accumulation).
 - **Single-step x₀ coordinate-space diffusion teacher**: completes a scene in one forward pass directly in coordinate space at 209 ms/frame (the SMC paper's efficiency contribution; it does not claim a scaffold-free accuracy win over baselines — that analysis lives in the RA-L paper below).
 
@@ -86,7 +86,7 @@ LiDAR vs DA2 (Depth Anything V2 monocular pseudo-LiDAR):
 | Parameters | 7.1 M |
 | Inference  | 1.6 ms / frame |
 | Decoded points | 8,000 |
-| Latent | 32 tokens × 1024-d |
+| Latent | 32 tokens × 32-d (1,024-d total) |
 
 ---
 
@@ -97,7 +97,7 @@ LiDAR vs DA2 (Depth Anything V2 monocular pseudo-LiDAR):
 | `teacher_v2gt/best_model.pth`                 | 542 MB | Diffusion teacher trained with GT-coordinate scaffold on v2 GT, 30 ep      | SMC + RA-L Tables I, IV, V, VI |
 | `teacher_v2gt_ft/best_model.pth` *(NEW)*      | 470 MB | Mixed-scaffold fine-tune (3 ep, 45 min, 30 % LIDAR-cropped scaffold mix), seed 42 reference; 6-seed evaluation in RA-L paper | RA-L Table III (the 0.968 m² row) |
 | `teacher_v1gt/best_model.pth`                 | 542 MB | Same architecture trained on v1 (standard) GT — for the 16× refinement comparison | SMC + RA-L Table I             |
-| `vae_v3/best_point_vae.pth`                   | ~85 MB | Multi-token Gaussian VAE (7.1 M params, 32 × 1024-d latent, 8,000 pts) — produced by `models/point_cloud_vae.py` (class `PointCloudVAE`), **not** the deprecated `point_cloud_vae_v2.py` FoldingNet design | SMC §III-B, Table II           |
+| `vae_v3/best_point_vae.pth`                   | ~85 MB | Multi-token Gaussian VAE (7.1 M params, 1,024-d latent (32 tokens × 32-d), 8,000 pts) — produced by `models/point_cloud_vae.py` (class `PointCloudVAE`), **not** the deprecated `point_cloud_vae_v2.py` FoldingNet design | SMC §III-B, Table II           |
 
 All four share the frozen Sonata / PTv3 encoder (108 M params), loaded from the upstream Sonata release; only the small denoiser / VAE heads are trained.