Yusin commited on
Commit
57a9f4c
·
verified ·
1 Parent(s): 61ec6b6

ReTokenGS-stream: epoch 29 of the streaming fine-tune

Browse files
Files changed (2) hide show
  1. README.md +31 -0
  2. model.safetensors +3 -0
README.md ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ tags: [gaussian-splatting, 3d-reconstruction, novel-view-synthesis, feed-forward, streaming]
4
+ datasets: [DL3DV/DL3DV-10K]
5
+ ---
6
+
7
+ # ReTokenGS — stream
8
+
9
+ Recurrent Gaussian Tokens with **uncertainty-driven births** and **streaming input views**, built on [LocusGS](https://github.com/leo-frank/LocusGS) / [TokenGS](https://github.com/nv-tlabs/TokenGS).
10
+
11
+ - **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.
12
+ - **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.
13
+ - **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.
14
+
15
+ The model works with posed views at 256×256. `model.safetensors` is epoch 29 of the streaming fine-tune.
16
+
17
+ | Official DL3DV benchmark (140 scenes, 4 views, 256×256, 32 rounds) | PSNR | LPIPS | SSIM |
18
+ |---|---|---|---|
19
+ | ReTokenGS-grad ([`Yusin/ReTokenGS-grad`](https://huggingface.co/Yusin/ReTokenGS-grad)) | 24.733 | 0.2664 | 0.7916 |
20
+ | ReTokenGS-dens, 4096 tokens ([`Yusin/ReTokenGS-dens`](https://huggingface.co/Yusin/ReTokenGS-dens)) | 25.091 | 0.2442 | 0.8031 |
21
+ | **ReTokenGS-stream** | **22.856** | **0.3881** | **0.7108** |
22
+
23
+ | Streaming (same 140 scenes, 6-view split, PSNR) | 16 rounds | 24 rounds | 32 rounds |
24
+ |---|---|---|---|
25
+ | all 6 views from round 1 | 23.421 | – | 24.074 |
26
+ | views 1–4 only | – | 22.817 | – |
27
+ | views 1–4, views 5–6 join at round 17 | – | 23.289 | 23.675 |
28
+
29
+ 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.
30
+
31
+ Code, evaluation and training commands: GitHub [`Yusin2Chen/TokenGS-Loop`, branch `stream`](https://github.com/Yusin2Chen/TokenGS-Loop/tree/stream). License: Apache-2.0 (derived from LocusGS / TokenGS).
model.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dc0ce5bd9e7742f0ce581b6a46c470141d2a7f704bf7fe7f5cd02c10f9b09121
3
+ size 705320996