| license: apache-2.0 | |
| tags: [gaussian-splatting, 3d-reconstruction, novel-view-synthesis, feed-forward] | |
| datasets: [DL3DV/DL3DV-10K] | |
| # ReTokenGS — dens | |
| Recurrent Gaussian Tokens with **on-demand densification**, built on [LocusGS](https://github.com/leo-frank/LocusGS) / [TokenGS](https://github.com/nv-tlabs/TokenGS): the `grad` model (one weight-shared decoder layer × 12 rounds, per-token render-gradient feedback) plus a split step after every round driven by the same gradient — tokens whose Gaussians receive large positional gradients split into opacity-corrected children, so the token budget grows per scene from 1024 and keeps improving up to a 4096 cap that was never used in training. 4 posed input views, 256×256. No error cross-attention. | |
| `model.safetensors` is the average of the last five training checkpoints (epochs 44–64). | |
| | Model (official DL3DV benchmark: 140 scenes, 4 views, 256×256) | Tokens at R12 | PSNR | LPIPS | SSIM | | |
| |---|---|---|---|---| | |
| | LocusGS stage 1, released `base` weights (same evaluation) | 1024 | 24.154 | 0.2766 | 0.7805 | | |
| | ReTokenGS-grad ([`Yusin/ReTokenGS-grad`](https://huggingface.co/Yusin/ReTokenGS-grad)) | 1024 | 24.733 | 0.2664 | 0.7916 | | |
| | ReTokenGS-dens, training threshold τ = 5.4e-5 | ≈ 1650 | 24.867 | 0.2611 | 0.7948 | | |
| | **ReTokenGS-dens, `DENS_TAU=3.4e-6 DENS_MAX_TOKENS=4096`** | 4096 | **25.091** | **0.2442** | **0.8031** | | |
| Code, evaluation and training commands: GitHub [`Yusin2Chen/TokenGS-Loop`, branch `dens`](https://github.com/Yusin2Chen/TokenGS-Loop/tree/dens). 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. | |