TokenGS-Loop

Feed-forward 3D Gaussian splatting from 4 posed views at 256×256, built on LocusGS / TokenGS. The stack of 12 independent LocusGS decoder layers is replaced by one decoder layer unrolled 12 times with shared weights. Three additions make this work:

• round-conditioned FiLM gates;
• render-error feedback from the input views;
• stop-gradient between the error-free first round and the refinement rounds.

file	branch	tokens	PSNR	LPIPS	SSIM
errenc-sg/model.safetensors (epoch 26)	errenc-sg	1024	23.706	0.3043	0.7586
split-clone/model.safetensors (epoch 17)	split-clone: error-guided token splitting, opacity-corrected children	364 → 1024	23.301	0.3401	0.7387
LocusGS stage 1, released base weights (reference, same evaluation)	–	1024	24.154	0.2766	0.7805

All rows use the official DL3DV benchmark: 140 scenes, 4 input views, 256×256, all 50 target frames.

Code: GitHub Yusin2Chen/TokenGS-Loop, branches errenc-sg and split-clone. Each branch README describes the method, evaluation and training commands. The checkpoints contain the weights of decoder blocks 1–11 inherited from initialization; only block 0 is executed.

License: Apache-2.0 (derived from LocusGS / TokenGS).