ReTokenGS β€” stream

Recurrent Gaussian Tokens with uncertainty-driven births and streaming input views, built on LocusGS / TokenGS.

  • The loop. One weight-shared decoder layer runs as a loop. Every round, each token reads the rendering gradient of its own Gaussians (G3R) and takes a line-searched latent gradient step.
  • Births. The tokens whose learned uncertainty head predicts the largest error split, and the number of splits per round is the predicted error itself. The loop starts at 256 tokens and grows to at most 1024, with 32 Gaussians per token.
  • Streaming. Input views can join at any round. An arriving view is encoded on arrival, and all tokens keep iterating on every view seen so far.

The model works with posed views at 256Γ—256. model.safetensors is epoch 29 of the streaming fine-tune.

Official DL3DV benchmark (140 scenes, 4 views, 256Γ—256, 32 rounds) PSNR LPIPS SSIM
ReTokenGS-grad (Yusin/ReTokenGS-grad) 24.733 0.2664 0.7916
ReTokenGS-dens, 4096 tokens (Yusin/ReTokenGS-dens) 25.091 0.2442 0.8031
ReTokenGS-stream 22.856 0.3881 0.7108
Streaming (same 140 scenes, 6-view split, PSNR) 16 rounds 24 rounds 32 rounds
all 6 views from round 1 23.421 – 24.074
views 1–4 only – 22.817 –
views 1–4, views 5–6 join at round 17 – 23.289 23.675

In the standard 4-view setting this model is behind grad and dens. Its contribution is the birth rule and the streaming mechanism: the late views add +0.47 dB over the views-1–4 reference at the same number of rounds.

Code, evaluation and training commands: GitHub Yusin2Chen/TokenGS-Loop, branch stream. License: Apache-2.0 (derived from LocusGS / TokenGS).

Downloads last month

-

Downloads are not tracked for this model. How to track
Safetensors
Model size
0.2B params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Dataset used to train Yusin/ReTokenGS-stream