Monocular 3D Gaussian Splatting SLAM suffers from critical limitations in time efficiency, geometric accuracy, and multi-view consistency. These issues stem from the time-consuming Train-from-Scratch optimization and the lack of inter-frame scale consistency from single-frame geometry priors. We contend that a feed-forward paradigm, leveraging multi-frame context to predict Gaussian attributes directly, is crucial for addressing these challenges. We present Flash-Mono, a system composed of three core modules: a feed-forward prediction frontend, a 2D Gaussian Splatting mapping backend, and an efficient hidden-state-based loop closure module. We trained a recurrent feed-forward frontend model that progressively aggregates multi-frame visual features into a hidden state via cross attention and jointly predicts camera poses and per-pixel Gaussian properties. By directly predicting Gaussian attributes, our method bypasses the burdensome per-frame optimization required in optimization-based GS-SLAM, achieving a 10x speedup while ensuring high-quality rendering. The power of our recurrent architecture extends beyond efficient prediction. The hidden states act as compact submap descriptors, facilitating efficient loop closure and global Sim(3) optimization to mitigate the long-standing challenge of drift. For enhanced geometric fidelity, we replace conventional 3D Gaussian ellipsoids with 2D Gaussian surfels. Extensive experiments demonstrate that Flash-Mono achieves state-of-the-art performance in both tracking and mapping quality, highlighting its potential for embodied perception and real-time reconstruction applications. Project page: https://victkk.github.io/flash-mono.
Figures & tables
Figure 1: Our Results for Reconstruction and Rendering & Tracking & Speed Metrics. Our method reconstructs high-quality Gaussian maps in complex scenes with multiple rooms and varying lighting conditions. The right-side radar chart shows our rendering quality (PSNR, SSIM, LPIPS) and trajectory tracking accuracy (ATE), with reciprocals of LPIPS, ATE, and Depth L1 plotted for clarity. Our method outperforms others in both rendering quality and trajectory accuracy, offering a 10x speedup over contemporary monocular GS-SLAM methods.
Figure 2: Pipeline. For each new frame, our recurrent model jointly infers the camera pose and per-pixel 2DGS attributes conditioned on a hidden state. The hidden state is updated simultaneously. To avoid catastrophic forgetting, the stream is partitioned into submaps. The hidden state is reinitialized for each submap. Past hidden states are cached in the Bag of Hidden States. Upon loop detection, i.e., revisiting a location, we perform a single forward pass on the loop frame conditioned on the past hidden state to relocalize the current frame in the past submap. A following pose graph optimization is then performed to correct the full trajectory. In the backend, per-frame 2DGS attributes prediction is voxelized, merged, and refined to build a global 2DGS map.
Figure 3: Qualitative Rendering Results.
ATE [cm] ↓
ScanNetV1
BundleFusion
0054
0059
0106
0169
0233
0465
apt0
apt2
copyroom
office0
office2
ORB-SLAM3
243.26
90.67
178.13
60.15
25.01
181.86
87.37
265.64
27.60
116.33
49.33
DROID-SLAM
161.22
69.92
89.11
28.26
74.01
117.27
89.38
148.04
19.71
31.41
73.91
MonoGS
70.19
97.24
150.89
191.98
62.45
113.19
122.59
142.54
53.41
62.67
127.02
DepthGS
192.18
93.69
140.19
205.92
81.90
121.01
67.52
119.74
14.59
40.42
16.05
S3PO-GS
69.36
16.52
26.15
87.04
27.09
96.35
92.49
97.90
21.88
64.22
69.88
Table 1: ATE RMSE (cm) on ScanNetV1 and BundleFusion datasets. Lower is better. We mark the first and second best results.
Method
Metric
ScanNetV1
BundleFusion
0054
0059
0106
0169
0233
0465
FPS ↑
apt0
apt2
copyroom
office0
office2
FPS ↑
MonoGS
SSIM ↑
0.80
0.74
0.72
0.77
0.68
0.59
0.69
0.70
0.39
0.70
0.52
0.60
1.00
LPIPS ↓
0.61
0.60
0.54
0.66
0.67
0.74
0.67
0.82
0.63
0.78
0.71
PSNR ↑
19.24
16.54
16.09
18.86
17.65
14.52
13.68
11.50
14.37
13.38
13.96
DepthGS
SSIM ↑
0.31
0.32
0.34
0.42
0.36
0.26
1.57
0.38
0.41
0.58
0.56
0.58
1.28
LPIPS ↓
0.79
0.78
0.78
0.73
0.84
0.81
0.67
0.69
0.51
0.62
0.63
Table 2: Mapping quality on ScanNetV1 and BundleFusion . Higher is better for SSIM/PSNR, lower is better for LPIPS. We mark the first and second best results.
Figure 4: Qualitative Analysis on Rendered Depth.
ATE RMSE [m] ↓
00
05
06
07
08
28
Ours
12.85
16.58
9.93
12.08
45.25
16.75
S3PO-GS
32.49
34.76
16.43
fail
64.74
23.64
Table 3: ATE RMSE (m) on KITTI Odometry . Lower is better.
Method
Metric
00
05
06
07
08
28
S3PO-GS
PSNR ↑
16.65
15.64
13.55
fail
17.25
15.30
SSIM ↑
0.5409
0.5320
0.4726
fail
0.5912
0.5053
LPIPS ↓
0.6254
0.6352
0.7241
fail
0.4626
0.6131
Ours
PSNR ↑
17.41
17.01
15.13
17.89
16.12
17.47
SSIM ↑
0.6584
0.6278
0.5922
0.6036
0.6221
0.5633
LPIPS ↓
0.5358
0.4871
0.5333
0.4854
0.4710
0.4581
Table 4: Rendering quality on KITTI Odometry . Higher is better for PSNR/SSIM, lower is better for LPIPS.
Figure 9
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Depth L1 [m] ↓
ScanNetV1
BundleFusion
0054
0059
0106
0169
0233
0465
apt0
apt2
copyroom
office0
office2
MonoGS
1.0559
1.2713
1.4149
1.563
0.8866
0.8163
0.9626
1.2832
1.1786
1.1547
1.2588
DepthGS
0.4515
0.6585
0.5227
0.4818
0.4051
0.4685
0.3673
0.29
0.1259
0.1841
0.2068
S3PO-GS
0.5805
0.3482
0.5535
0.6612
0.2769
0.8873
0.7191
0.9869
0.4116
1.0105
0.8539
Ours
0.16
0.23
0.51
0.17
0.35
0.44
0.33
0.3455
0.11
0.11
0.18
Appendix
Table 6: Depth L1 Error (m) on ScanNetV1 and BundleFusion datasets. Lower is better. We mark the best results.
Figure 6: Qualitative Analysis on Estimated Trajectory
Figure 7: Qualitative Analysis on reconstructed ScanNet scene 0054. All baselines failed to reconstruct the scene.
Component
Total Parameters
Encoder
303.1 M
Decoder
380.8 M
Heads & Tokens
111.8 M
Total
795.7 M
Appendix
Table 7: Detailed breakdown of Flash-Mono model parameters.
Figure 8: CUDA Graph optimization
Thread
Module
Time (ms)
Note
Frontend
Feedforward Inference
62
Per frame
Loop Closure Detection
3
Per frame
Total (Per Frame)
65
Loop Frame Feedforward
62
Per loop closure
PGO (Sim3 Optimization)
32
Per loop closure
Backend
Merge & Voxelization
0.5
Per frame
Appendix
Table 8: Runtime breakdown of Flash-Mono. The system runs in parallel threads, with the Backend being the primary bottleneck. Loop closure operations are sparse events.
Method
fr1/desk
fr2/xyz
fr3/office
MonoGS
26.64k
43.59k
35.24k
Photo-SLAM
40.00k
0.10m
81.16k
SplaTAM
0.96m
6.36m
0.79m
Gaussian-SLAM
0.76m
0.69m
1.47m
GS-ICP-SLAM
0.53m
1.91m
2.09m
Ours
0.63m
0.98m
0.61m
Appendix
Table 9: Quantitative comparison of the total Gaussian count on the TUM dataset. Our method maintains a balance between map density and compactness.
Figure 9: Case Study: Robust Relocalization Under Environmental Changes. The model generates a hidden state from 8 context views captured at night (curtains closed, empty chair). When presented with a new observation from the same location but under drastically different conditions (daytime, curtains open, person sitting), the feed-forward model successfully relocalizes and reconstructs accurate geometry. This demonstrates the hidden state’s potential for life-long mapping scenarios where environments undergo temporal changes.
State Key Laboratory of Industrial Control Technology, College of Control Science and Engineering, Zhejiang University, Hangzhou, China, 310027. · Key Laboratory of Collaborative Sensing and Autonomous Unmanned Systems of Zhejiang Province, Hangzhou, China, 310027.