Scene chronology aims to reconstruct a scene across visits separated by hours or days, during which objects may change in position, presence or shape. Such reconstruction must retain every earlier state, yet each revisit often provides only a few images. Storing a complete model for each visit duplicates the unchanged scene, whereas updating a single shared model alters earlier states. We present ReScene, which stores each visit as an immutable compact update, so that any earlier visit can be recovered exactly by applying the stored updates in order, without retaining past images. To localise each update, ReScene attributes the disagreement between the preceding reconstruction and the new images to individual Gaussians through their compositing weights. This attribution guides the suppression of outdated Gaussians and the refinement of persistent ones, while newly visible surfaces are initialised from learned multi-view depth. We also introduce Multi-State, a benchmark of non-rigid deformation and partial material removal, including synthetic scenes with ground-truth geometry. ReScene achieves quality higher than sequential 3DGS checkpointing with 2.5--4.8× less uncompressed history storage. At comparable PSNR, it also reduces foreground depth error by 3.2× relative to sequential ChronoGS, showing that accurate rendering does not necessarily imply accurate geometry. Code and benchmark are available at https://waychin-weiqin.github.io/ReScene/.
Figures & tables
Figure 1: Sample data and history storage on Multi-State. Top: deformation of a cloth (synthetic) and material removal from a tomato (real). Bottom: quality and history storage comparison on real Multi-State. Each point averages held-out PSNR over all visits observed so far, each recovered after the latest update. Marker area denotes history storage. ReScene retains the quality of earlier visits at a fraction of the storage required by checkpointing (Seq. 3DGS).
Figure 2: Overview of ReScene. For each visit t , the preceding reconstruction Gt−1 is rendered from the viewpoint of the new images, and the disagreement between the renderings and the images is attributed to individual Gaussians as change support Ci and stable support Si . A visit-specific MLP ft maps the features zit of each Gaussian to opacity logits, which suppress Gaussians of moved or removed surfaces and determine the opacity of births. Gaussian offsets refine existing Gaussians, whereas births are initialised from depth seeds, and densified by splitting. The resulting update Δt produces Gt , whose rendering after the visit-specific appearance correction et is compared with the observed images to optimise Eq. ( 5 ).
Multi-State (Real)
Multi-State (Synthetic)
WAT
ChronoScene
Method
PSNR
SSIM
Storage
PSNR
SSIM
Storage
PSNR
SSIM
Storage
PSNR
SSIM
Storage
CL-Splats
21.03
0.826
91.2
29.34
0.942
220.5
18.10
0.731
174.7
18.74
0.605
1581.5
Seq. ChronoGS (arrival)
24.92
0.863
108.6
32.07
0.954
108.7
27.42
0.891
368.0
20.69
0.613
706.7
Seq. ChronoGS (recall)
18.90
0.774
25.81
0.907
18.75
0.726
16.53
0.467
Seq. LTGS
21.87
0.844
304.7
28.81
0.938
995.0
23.24
0.834
873.2
19.41
0.617
5664.1
Seq. 3DGS
24.02
0.861
410.5
33.24
0.956
1673.3
25.67
0.863
1816.1
19.57
0.577
9618.7
Table 1: Reconstruction quality and history storage across three benchmarks. Quality is reported as PSNR (dB) and SSIM (higher is better for both), averaged equally across scenes. Seq. ChronoGS queries its latest shared model, so its arrival and final-recall quality are reported separately; the other methods preserve arrival quality through stored histories. Storage is the mean uncompressed model and history size per scene (MiB, lower is better). Bold marks the best final-recall quality and the lowest storage in each column.
Method
Robot
Teddy
Plant
Cloth
Mean
Ours
30.22
7.10
121.80
19.23
44.59
Seq. LTGS
197.63
29.90
101.21
132.54
115.32
Seq. ChronoGS (arrival)
153.10
27.54
144.45
251.67
144.19
Seq. ChronoGS (recall)
1119.22
46.48
232.49
247.18
411.34
Table 2: Geometric error of foreground objects on synthetic Multi-State. We report the mean absolute error (lower is better) of the rendered expected depth along the camera axis, evaluated using ground-truth foreground masks and rendered opacity ≥0.5 . The reported error is averaged over all visits, including the initial one. The lowest error in each column is shown in bold.
Figure 3: Final recall and arrival on real Multi-State. Each scene pairs an earlier visit recovered after the final update (upper row) with a later visit at arrival (lower row). Our method and Seq. LTGS recover earlier states exactly; Seq. ChronoGS alters them. Cameras are matched within rows; overlays report whole-image PSNR.
Real
Synthetic
Change masks
Attribution
PSNR ↑
SSIM ↑
PSNR ↑
SSIM ↑
LTGS
Compositing
23.33
0.855
30.60
0.947
CL-Splats
Voting
22.09
0.840
31.25
0.953
CL-Splats
Compositing
24.79
0.870
31.63
0.954
Ours
Compositing
25.34
0.872
32.50
0.957
Table 3: Change localisation and attribution on the Multi-State dataset . We compare our proposed mechanism with alternatives derived in CL-Splat [ 1 ] and LTGS [ 7 ] .
Figure 4: View-count sensitivity on real Multi-State. Arrival and final-recall PSNR across four scenes, including initialisation. Differences in the text use unrounded means. The horizontal axis gives the number of current images at each revisit.
Synthetic
Real
Removed
PSNR ↑
SSIM ↑
MiB ↓
PSNR ↑
SSIM ↑
MiB ↓
Births
27.82
0.929
271.5
23.44
0.851
120.2
Suppression
28.64
0.933
341.8
22.76
0.846
155.7
Refinement
30.67
0.952
186.5
22.14
0.837
126.8
Splitting
31.49
0.950
308.9
24.99
0.866
141.4
None (full)
32.11
0.955
352.0
25.16
0.870
163.6
Table 4: Component ablations on Multi-State. Each row removes one component from our proposed method. Reported results exclude the initial visit, and are averaged across all scenes.
Limit
Synthetic
Real
PSNR ↑
PSNR FG↑
PSNR BG↑
SSIM ↑
LPIPS ↓
Storage ↓
PSNR ↑
SSIM ↑
LPIPS ↓
Storage ↓
5,000
31.16
23.00
35.14
0.948
0.081
286.9
24.50
0.859
0.359
124.1
10,000
31.54
23.63
35.23
0.950
0.077
302.2
24.76
0.863
0.352
128.0
20,000
31.91
24.27
35.29
0.953
0.073
323.4
24.91
0.865
0.345
136.3
30,000
32.06
24.54
35.31
0.954
0.072
339.4
25.02
0.867
0.340
144.6
40,000
32.11
24.65
35.31
0.955
0.071
346.4
25.15
0.869
0.335
152.4
Table S1 : Birth limits on Multi-State. Each limit bounds the total number of active births per visit, including children created by splitting. Quality is averaged over later visits, excluding initialisation, and equally across scenes. On the synthetic scenes, PSNR FG and PSNR BG denote foreground and background PSNR computed using ground-truth object masks. All other quality metrics are computed on whole images. Storage is the mean uncompressed history size per scene in MiB. The shaded row is the default limit. Limits above 64,306 never become active.
Figure S1 : Changing states in Multi-State. (a) First, intermediate and final states of Robot, Teddy, Plant and Cloth, rendered from a fixed held-out camera. (b) Can, Tissue, Tomato and Gimbal at visits 0–3, shown using captured held-out images. The real viewpoints are approximately matched across visits, rather than identical.
Scene
States
Train Views
Test Views
Can
4
337
45
Tissue
4
461
59
Tomato
4
462
59
Gimbal
4
398
51
Table S2 : Real capture collection and evaluation subset.
Scene
Variant
PSNR ↑
SSIM ↑
MiB ↓
Robot
Full
35.71
0.9746
386.3
No births
26.35
0.9420
285.6
No suppression
28.54
0.9471
382.4
No refinement
32.65
0.9699
204.2
No splitting
34.95
0.9729
334.7
Teddy
Full
31.84
0.9501
306.5
Table S3 : Scene-wise component ablations on synthetic Multi-State. PSNR and SSIM exclude initialisation. Storage is the complete uncompressed history in MiB. Bold marks the best quality per scene.
Scene
Variant
PSNR ↑
SSIM ↑
MiB ↓
Can
Full
23.59
0.8633
160.7
No births
22.20
0.8494
118.7
No suppression
21.84
0.8427
151.9
No refinement
22.06
0.8401
124.1
No splitting
23.40
0.8602
139.2
Tissue
Full
23.82
0.8548
167.9
Table S4 : Scene-wise component ablations on real Multi-State. PSNR and SSIM exclude initialisation. Storage is the complete uncompressed history in MiB. Bold marks the best quality per scene.
Method
Views
All A
All R
Later A
Later R
SSIM A
SSIM R
LPIPS A
LPIPS R
Ours
3
21.26
21.26
19.64
19.64
0.8103
0.8103
0.3824
0.3824
Ours
6
22.75
22.75
21.63
21.63
0.8356
0.8356
0.3584
0.3584
Ours
12
24.11
24.11
23.44
23.44
0.8574
0.8574
0.3387
0.3387
Ours
24
25.40
25.40
25.17
25.17
0.8732
0.8732
0.3221
0.3221
Ours
36
25.67
25.67
25.52
25.52
0.8770
0.8770
0.3184
0.3184
Ours
48
25.91
25.91
25.85
25.85
0.8795
0.8795
0.3165
0.3165
Table S5 : Equal-scene quality across four real Multi-State scenes. A and R denote arrival and final recall, respectively. All includes initial model while Later excludes it. SSIM and LPIPS refer to all visits (including the initial model).
Views
Ours
Seq. ChronoGS
Seq. 3DGS
CL-Splats
3
115.6/0.43
98.7/0.49
344.1/0.27
95.0/0.32
6
135.7/0.44
107.2/0.49
358.0/0.26
90.3/0.33
12
153.5/0.45
112.8/0.49
383.9/0.26
89.9/0.34
24
163.6/0.46
108.6/0.48
410.5/0.27
91.2/0.33
36
168.2/0.46
104.9/0.47
430.6/0.27
92.0/0.32
48
167.3/0.46
100.4/0.46
440.4/0.28
91.8/0.32
Table S6 : Storage and update-training cost by view count. Each cell reports mean uncompressed history storage per scene (MiB) / total update-training time across four scenes (hours). Time excludes initial fitting, preparation, evaluation and failed attempts.
Figure S2 : Arrival and final recall on synthetic Multi-State. Cloth, Robot and Teddy. Each scene occupies two rows with identical ground truth and camera. Here, t denotes the reconstructed visit and T the final update. Overlays show whole-image PSNR (dB).
Figure S3 : Arrival and final recall on real Multi-State. Can, Tomato and Gimbal. Each scene occupies two rows with identical ground truth and camera. Here, t denotes the reconstructed visit and T the final update. Overlays show whole-image PSNR (dB).
Figure S4 : Arrival and final recall on WAT. Living Room, Dyson and Mac. Each scene occupies two rows with identical ground truth and camera. Here, t denotes the reconstructed visit and T the final update. Overlays show whole-image PSNR (dB).
Figure S5 : Arrival and final recall on ChronoScene. Aerial2 and Street2. Each scene occupies two rows with identical ground truth and camera. Here, t denotes the reconstructed visit and T the final update. Overlays show whole-image PSNR (dB).
Figure S6 : Failure case on Plant. Comparison at visit 9 from held-out view 0. Both sequential baselines achieve higher whole-image PSNR than our method in this example, illustrating a limitation when reconstructing thin structures with substantial occlusion.
Reconstructing environments where parts of the scene change between captured image sets poses a challenge for 3D scene reconstruction. We present ChronoFuseGS, a multi-temporal Gaussian Splatting approach that addresses this issue by taking multiple separately trained Gaussian Splatting models, each representing a distinct timestep and partially overlapping in geographic coverage, and merging them into a single combined model. By allowing Gaussians from one timestep to contribute to the reconstruction at other timesteps, our approach leverages data across all captured timesteps to refine persistent parts of the scene. The model supports incremental extension, allowing new timesteps to be added while preserving the existing merged reconstruction. It encodes, for each Gaussian primitive, at which timesteps it contributes to the reconstruction. To support visual exploration of the reconstructed scene, we present a change-aware visualization approach that highlights the parts of the scene that have changed across a user-defined time selection, while preserving the color of persistent parts. Since the persistence encoding operates at the Gaussian primitive level, changes are visualized at sub-object granularity rather than being limited to object-level changes. We evaluate our approach on a real-world outdoor dataset of a flood management area, captured over 7 months across eight recording days and covering seasonal vegetation changes, snow cover, and flooding events, which we make publicly available. Our results demonstrate that the combined model consistently outperforms individually trained single-timestep models in novel-view synthesis quality, recovers structural details absent in the individual reconstructions, and reliably highlights changes in fine details and sub-parts of objects and natural structures.
We introduce ReconSplat, a feed-forward model for 3D scene reconstruction that aims to address the longstanding trade-off between plausible view generation for unobserved regions and geometric consistency, providing both geometrically aligned novel views and sharp depth estimates. Our approach builds on 3D Gaussian splatting (3DGS) as an intermediate differentiable scene representation and integrates it with a multi-view latent diffusion model (MV-LDM) trained to act simultaneously as a refiner and an inpainter for appearance and scene geometry. We enforce geometric consistency by guiding the diffusion process with variational 3D latent features for appearance and geometry, encoded by the feed-forward 3DGS representation and rasterized to 2D latent space. ReconSplat produces both photorealistic novel views and accurate depth maps on real-world benchmarks, RealEstate10K and DL3DV-10K, outperforming existing methods in challenging extrapolation setups. Notably, ReconSplat allows the extrapolation of unseen and challenging viewpoints jointly with coherent and precise scene geometry.
Giuseppe Stracquadanio, Kevin Raj, Julia Grabinski +1
Streaming 3D reconstruction models keep past video frames in a recurrent state, fast weights from test-time training, or a key-value (KV) cache. Each written frame shifts the coordinate frame of later cameras or occupies GPU memory. Training-free rules decide what to write from a signal of the current frame. How much they write accounts for their mean accuracy: spending the update budget of a gate at uniformly spaced frames, with no signal, matches nine published gates, and the budget bounds the drift of the state wherever the writes fall. We propose SceneBound, a training-free rule that writes a frame only when its predicted camera pose lies at least a fixed radius from every stored keyframe. Its keyframe count stays within a constant factor of the covering number of the visited poses, the fewest views that keep a keyframe near every camera, so the count grows with the places visited and stays flat when the camera returns. In large scenes, SceneBound closes a submap once decoding its first frame again reveals drift. SceneBound uses one radius across four released models and three memory types. In the recurrent state of CUT3R and the KV cache of StreamVGGT, it achieves the lowest mean trajectory error among the published training-free rules we evaluate, in one pass and over nine passes of a scene, where its keyframes grow by 3% to 6%.
Ziming Wang, Chuyue Gong, Chutian Wang +4
The Hong Kong University of Science and Technology · Sun Yat-sen University · Monash University +2