ReTokenGS — grad
Recurrent Gaussian Tokens for query-based feed-forward 3D Gaussian Splatting, built on LocusGS / TokenGS: the 12 independent decoder layers are replaced by one weight-shared layer unrolled 12 times, and after every round each token reads the gradient of the input-view rendering loss with respect to its own Gaussians (after G3R); no error cross-attention. 4 posed input views, 256×256, 1024 tokens × 64 Gaussians.
model.safetensors is the average of the last six training checkpoints (epochs 44–66). It also contains the weights of decoder blocks 1–11 inherited from initialization; only block 0 is executed.
| Model (official DL3DV benchmark: 140 scenes, 4 views, 256×256) | Tokens | PSNR | LPIPS | SSIM |
|---|---|---|---|---|
LocusGS stage 1, released base weights (same evaluation) |
1024 | 24.154 | 0.2766 | 0.7805 |
| ReTokenGS-grad (this file) | 1024 | 24.733 | 0.2664 | 0.7916 |
ReTokenGS-dens, 4096 tokens on demand (Yusin/ReTokenGS-dens) |
≤ 4096 | 25.091 | 0.2442 | 0.8031 |
Code, evaluation and training commands: GitHub Yusin2Chen/TokenGS-Loop, branch grad. License: Apache-2.0 (derived from LocusGS / TokenGS).
The numbers were produced with the released code, which has no error cross-attention; the loop_err_* tensors in this file are unused and ignored at load time.