SWinGS: Sliding Windows for

Dynamic 3D Gaussian Splatting

Richard Shaw1 , Michal Nazarczuk1 , Jifei Song1 , Arthur Moreau1 , Sibi

Catley-Chandar1,2 , Helisa Dhamo1 , and Eduardo Pérez-Pellitero1
1
Huawei Noah’s Ark Lab
2
Queen Mary University Of London
arXiv:2312.13308v2 [cs.CV] 18 Jul 2024

MixVoxels Ours HyperReel Ours

MixVoxels Ours HyperReel Ours

Fig. 1: Left: SWinGS achieves sharper dynamic 3D scene reconstruction in part thanks
to a sliding window canonical space that reduces the complexity of the 3D motion
estimation. Right: Our dynamic real-time viewer allows users to explore the scene.

Abstract. Novel view synthesis has shown rapid progress recently, with
methods capable of producing increasingly photorealistic results. 3D
Gaussian Splatting has emerged as a promising method, producing high-
quality renderings of scenes and enabling interactive viewing at real-time
frame rates. However, it is limited to static scenes. In this work, we ex-
tend 3D Gaussian Splatting to reconstruct dynamic scenes. We model
a scene’s dynamics using dynamic MLPs, learning deformations from
temporally-local canonical representations to per-frame 3D Gaussians.
To disentangle static and dynamic regions, tuneable parameters weigh
each Gaussian’s respective MLP parameters, improving the dynamics
modelling of imbalanced scenes. We introduce a sliding window training
strategy that partitions the sequence into smaller manageable windows
to handle arbitrary length scenes while maintaining high rendering qual-
ity. We propose an adaptive sampling strategy to determine appropriate
window size hyperparameters based on the scene’s motion, balancing
training overhead with visual quality. Training a separate dynamic 3D
Gaussian model for each sliding window allows the canonical represen-
tation to change, enabling the reconstruction of scenes with significant
geometric changes. Temporal consistency is enforced using a fine-tuning
step with self-supervising consistency loss on randomly sampled novel
2 R. Shaw et al.

views. As a result, our method produces high-quality renderings of gen-

eral dynamic scenes with competitive quantitative performance, which
can be viewed in real-time in our dynamic interactive viewer.

1 Introduction
Photorealistic rendering and generally 3-dimensional (3D) imaging have received
significant attention in recent years, especially since the seminal work of Neural
Radiance Fields (NeRF) [37]. This is in part thanks to its impressive novel view
synthesis results, but also due to its appealing ease of use when coupled with off-
the-shelf structure-from-motion camera pose estimation [53]. NeRF’s key insight
is a fully differentiable volumetric rendering pipeline paired with learnable im-
plicit functions that model a view-dependent 3D radiance field. Dense coverage
of posed images of the scene provides then direct photometric supervision.
The original formulation of NeRF and most follow-ups [5,9,37,39,41] assume
static scenes and thus a fixed radiance field. Some have explored new paradigms
enabling dynamic reconstruction for radiance fields, including D-NeRF [22] and
Nerfies [43], optimising an additional continuous volumetric deformation field
that warps each observed point into a canonical NeRF. Such an approach has
been popular [12,13,28,32,33,44], and has also been used for dynamic human re-
construction [45, 74]. However, learning 3D deformation fields is inherently chal-
lenging, especially for large motions, with increased computational expense in
training and inference. Moreover, approaches that share a canonical space among
all frames struggle to maintain reconstruction quality for long sequences, obtain-
ing overly blurred results due to inaccurate deformations and limited represen-
tational capacity. Other methods avoid maintaining a canonical representation
and use explicit per-frame representations of the dynamic scene. Examples are
tri-plane extensions to 4D (x, y, z, t) [7,14,54,60] with plane decompositions [9] to
keep memory footprint under control. These approaches can suffer from a lack of
temporal consistency, especially as they generally are agnostic about the motion
of the scene. Grid-based methods that share a representation [46], also can suffer
from degradation due to lack of representational capacity in long sequences.
This paper proposes a new method that addresses open problems of the state-
of-the-art (SoTA) - see Fig. 1. Our method overview is shown in Fig. 2. Firstly,
we build upon 3D Gaussian Splatting (3DGS) [20] and adapt the 3D Gaussians
to be dynamic by allowing them to move. Our representation is thus explicit
and avoids expensive raymarching via fast rasterization. Secondly, we introduce
a novel paradigm for dynamic neural rendering with temporally-local canonical
spaces defined in a sliding window fashion. Each window’s length is adaptively
defined following the amount of scene motion to maintain high-quality recon-
struction. By limiting the scope of each canonical space, we can accurately track
3D displacements (i.e. they are generally smaller displacements) and prevent
intra-window flickering. Thirdly, we introduce tuneable MLPs (MLP with sev-
eral sets of weights governed by per-3D Gaussian blending weights) to estimate
displacements. This tackles scenes with static vs dynamic imbalance. By learn-
ing different “modes” of motion estimation, we can separate smoothly between
 SWinGS 3

static and dynamic regions with virtually no additional computational cost nor
any handcrafted heuristics. Lastly, temporal consistency loss computed on over-
lapping frames of neighbouring windows ensures consistency between windows,
i.e. avoids inter-window flickering. In summary, our main contributions are:
1. An adaptive sliding window approach that enables the reconstruction of
arbitrary length sequences whilst maintaining high rendering quality.
2. Temporally-local dynamic MLPs that model scene dynamics by learning
deformation fields from per-window canonical 3D Gaussians to each frame.
3. Learnable MLP tuning parameters tackle scene imbalance by learning differ-
ent motion modes; disentangling static canonical and dynamic 3D Gaussians.
4. A fine-tuning stage ensures temporal consistency throughout the sequence.

2 Related work
Non-NeRF dynamic reconstruction: A number of approaches prior to the
emergence of NeRF [37] tackled dynamic scene reconstruction. Such methods
typically relied on dense camera coverage for point tracking and reprojection [19],
or the presence of additional measurements, i.e. depth [40, 55]. Alternatively,
some were curated towards specific domains, e.g. car reconstruction in driving
scenarios [6,34]. Several recent works [3,29,46] followed the idea of Image-Based
Rendering [63] (direct reconstruction from neighbouring views). Others [30, 69]
utilize multiplane images [58, 76] with an additional temporal component.
NeRF-based reconstruction: NeRF [37] has achieved great success for recon-
structing static scenes with many works extending it to dynamic inputs. Several
methods [4, 67] model separate representations per time-step, disregarding the
temporal component of the input. D-NeRF [22] reconstruct the scene in a canon-
ical representation and model temporal variations with a deformation field. This
idea was developed upon by many works [12,13,28,32,33,44]. StreamRF [23] and
NeRFPlayer [56] use time-aware MLPs and a compact 3D grid at each time-step
for 4D field representation, reducing memory cost for longer videos. Other ap-
proaches [7,14,54,60] represent dynamic scenes with a space-time grid, with grid
decomposition to increase efficiency. DyNeRF [25] represents dynamic scenes by
extending NeRF with an additional time-variant latent code. HyperReel [2] uses
an efficient sampling network, modeling the scene around keyframes. MixVox-
els [61] represents dynamic and static components with separately processed
voxels. Several methods [18, 24, 64] rely on an underlying template mesh (e.g.
human). Some methods [8, 50, 75] aim to improve NeRF quality post rendering.
3D Gaussian Splatting Much of the development of neural rendering has
focused on accelerating inference [15,46,49,62,73]. Recently, 3D Gaussian Splat-
ting [20] made strides by modelling the scene with 3D Gaussians, which, when
combined with tile-based differentiable rasterization, achieves very fast render-
ing, yet preserves high-quality reconstruction. Luiten et al . [35] extend this to dy-
namic scenes with shared 3D Gaussians that are optimised frame-by-frame. How-
ever, their focus is more on tracking 3D Gaussian trajectories rather than maxi-
mizing final rendering quality. Concurrent methods to ours that extend 3D Gaus-
4 R. Shaw et al.

𝑆𝑙𝑖𝑑𝑖𝑛𝑔 𝑊𝑖𝑛𝑑𝑜𝑤 𝑆𝑎𝑚𝑝𝑙𝑖𝑛𝑔 𝐷𝑦𝑛𝑎𝑚𝑖𝑐 3𝐷 𝐺𝑎𝑢𝑠𝑠𝑖𝑎𝑛𝑠 𝑇𝑒𝑚𝑝𝑜𝑟𝑎𝑙 𝐶𝑜𝑛𝑠𝑖𝑠𝑡𝑒𝑛𝑐𝑦

(𝑝𝑒𝑟 − 𝑤𝑖𝑛𝑑𝑜𝑤)

𝛾(𝒙) ∆𝒙 𝑇𝑟𝑎𝑖𝑛 𝑣𝑖𝑒𝑤

𝛾(𝑡) ∆𝒓 𝑁𝑜𝑣𝑒𝑙 𝑣𝑖𝑒𝑤
∆𝒔

𝐶𝑜𝑛𝑠𝑖𝑠𝑡𝑒𝑛𝑐𝑦 𝐿𝑜𝑠𝑠
𝑂𝑝𝑡𝑖𝑐𝑎𝑙 𝐹𝑙𝑜𝑤

𝑇𝑖𝑚𝑒 𝑇𝑖𝑚𝑒

Fig. 2: Method. First, the sequence is partitioned into sliding windows based on
optical flow. Second, a dynamic 3DGS model is trained per window, where tunable
MLPs model the deformations. Blending parameters α weigh the MLP’s parameters
to focus on dynamic parts. Finally, each model is fine-tuned, enforcing inter-window
temporal consistency with consistency loss on sampled views for overlapping frames.

sian Splatting to focus on general dynamic scenes include [17, 26, 31, 57, 71, 72]
amongst others, while other works have focused specifically on dynamic human
reconstruction [16, 21, 27, 38, 42, 48, 65] and facial animation [11, 47, 52, 68, 70].

3 Method

3.1 Overview

We present our method to reconstruct and render novel views of general dynamic
scenes from multiple calibrated time-synchronized cameras. An overview of our
method is shown in Fig. 2. We build upon 3D Gaussian Splatting [20] for novel
view synthesis of static scenes, which we extend to scenes containing motion.
Our method can be separated into three main steps. First, given a dynamic
sequence, we split the sequence into separate shorter sliding windows of frames
for concurrent processing. We adaptively sample windows of varying lengths de-
pending on the amount of motion in the sequence. Each sliding window contains
an overlapping frame between adjacent windows, enabling temporal consistency
to be enforced throughout the sequence at a subsequent training stage. Parti-
tioning the sequence into smaller windows enables us to deal with sequences of
arbitrary length while maintaining high render quality.
Second, we train separate dynamic 3DGS models for each sliding window in
turn. We extend the static 3DGS method to model the dynamics by introducing
a tuneable MLP [36]. The MLP learns the deformation field from a canonical set
of 3D Gaussians for each frame in a window. Thus each window comprises an
independent temporally-local canonical 3D Gaussian representation and defor-
mation field. This enables us to handle significant geometric changes and/or if
new objects appear throughout the sequence, which can be challenging to model
with a single representation. A tuneable MLP weighting parameter α is learned
 SWinGS 5

for each 3D Gaussian to enable the MLP to focus on modelling the dynamic parts
of the scene, with the static parts encapsulated by the canonical representation.
Third, once a dynamic 3DGS model is trained for each window in the se-
quence, we apply a fine-tuning step to enforce temporal consistency throughout
the sequence. We fine-tune each 3DGS model sequentially and employ a self-
supervising temporal consistency loss on the overlapping frame renders between
neighbouring windows. This encourages the model of the current window to
produce similar renderings to the previous window. The result is a set of per-
frame Gaussian Splatting models enabling high-quality novel view renderings
of dynamic scenes with real-time interactive viewing capability. Our approach
enables us to overcome the limitations of training with long sequences and to
handle complex motions without exhibiting distracting temporal flickering.

3.2 Preliminary: 3D Gaussian Splatting

Our method is built upon 3D Gaussian Splatting (3DGS) [20]. 3DGS uses a 3D
Gaussian representation to model scenes as they are differentiable and can be
projected to 2D splats, enabling fast tile-based rasterization. The 3D Gaussians
are defined by 3D covariance matrix Σ in world space centered at the mean µ:

G(x) = e^{-\frac {1}{2}(x)^T \Sigma ^{-1} (x)}. (1)

Given a scaling matrix S and rotation matrix R, the corresponding covari-
ance matrix Σ of a 3D Gaussian can be written as Σ = RSS T RT . To represent
a scene, 3DGS optimizes 3D Gaussian positions x, covariance Σ (scaling S and
rotation R), opacity o, and colours, represented by spherical harmonic (SH)
coefficients, capturing view-dependent appearance. The optimization is inter-
leaved with Gaussian adaptive density control (ADC). The model is optimized
by rendering the learned Gaussians via a differentiable rasterizer, comparing the
resulting image Ir against the ground truth Igt , and minimizing the loss function:

\mathcal {L} = (1-\lambda )\mathcal {L}_1(I_r, I_{gt}) + \lambda \mathcal {L}_{\mathrm {SSIM}}(I_r,I_{gt}). (2)

3.3 Sliding-Window Processing

Using a single dynamic 3DGS representation to model an entire sequence be-

comes impractical for longer sequences, simply from a data processing stand-
point. Furthermore, representing a lengthy sequence with a single model per-
forms significantly worse than using multiple smaller fixed-sized segments, par-
ticularly if the scene has large motion or topological changes that cannot be
modeled easily with one canonical representation and deformation field (Fig. 4).
To address this, we use a sliding-window approach, separating the sequence into
smaller windows with overlapping frames. The window size is a hyperparameter
whose effect is explored in Section 4.1. Each window comprises an independent
dynamic 3DGS model, and we allow all 3D Gaussian parameters to change be-
tween windows, including the number of Gaussians, their positions, rotations,
6 R. Shaw et al.

scaling, colours and opacities. The advantage of this approach is that indepen-
dent models can be trained in parallel across multiple GPUs to speed up training.

Adaptive Window Sampling We propose an adaptive method for choosing

the appropriate sliding window sizes, balancing training overhead and model
size with performance and temporal consistency. Given an input sequence of
frame length Nf , we separate the sequence into smaller windows by adaptively
sampling windows of different lengths {Nw }, depending on the amount of motion
in the sequence. In high-motion areas, we aim to sample windows more frequently
(shorter windows), and in low-motion areas, less frequently (longer windows).
To do this, we leverage the magnitude of 2D optical flow from each cam-
era viewpoint. We employ a greedy algorithm to adaptively select the sizes of
windows, prior to training, based on the accumulated optical flow magnitude.
Specifically, we estimate per-frame optical flow f using a pre-trained RAFT [59]
model for each camera view j ∈ V and all frames in the sequence i ∈ Nf , and
compute the mean flow magnitude summed over each frame:

\hat {v}_i = \frac {1}{V} \sum ^{V}_j \sum ^{N_f-1}_i || \boldsymbol {f} (I^j_i, I^j_{i+1} ) ||^2_2 (3)

We iterate over each frame in the sequence with a greedy heuristic; spawning
a new window when the sum of mean flow v̂i exceeds a pre-defined threshold. This
ensures that each window contains a similar amount of movement, leading to a
balanced distribution of the total representational workload. Taking the average
across viewpoints makes this approach somewhat invariant to the number of
cameras, while placing a limit on the total flow stops an excessive amount of
movement within each window. Note, each sampled window overlaps with the
next window by a single frame, such that neighbouring windows share a common
image frame. This is to enable inter-window temporal consistency (section 3.6).

3.4 Dynamic 3D Gaussians

To extend static 3DGS to handle dynamic scenes, we introduce a temporally-
local dynamic MLP unique to each sliding window in the sequence (section 3.3).
Each time-dependent MLP learns a temporally-local deformation field, mapping
from a learned per-window canonical space to a set of 3D Gaussians for each
frame in the window. Each deformation field, represented by a small MLP Fθ
with weights θ, takes as input the normalized frame time t ∈ [0, 1] and 3D
Gaussian means x (normalized by the scene’s mean and standard deviation),
and outputs displacements to their positions ∆x, rotations ∆r and scaling ∆s:

\Delta \boldsymbol {x}(t), \Delta \boldsymbol {r}(t), \Delta \boldsymbol {s}(t) = \mathcal {F}_{\theta } (\gamma ({\boldsymbol {x}}), \gamma (t)) (4)
3 3+6m
where γ(.) denotes sinusoidal positional encoding γ : R → R , γ(x) =
(x, . . . , sin(2kπx), cos(2kπx), . . . ). We use small MLPs to reduce overfitting, set-
ting the number of frequency components m = 6, MLP depth D = 4 and width
W = 16, with two skip connections.
 SWinGS 7

3.5 Tunable Dynamic MLPs

Ideally, time-dependent MLPs model motion in the scene by associating the dy-
namic parts with the temporal input component γ(t) and thus learn to decouple
the scene into i) static canonical 3D Gaussians and ii) dynamic 3D Gaussians.
However, in scenes with an imbalance of static vs dynamic parts, e.g. Neural
3D Video [25] with mostly static backgrounds, the MLP struggles to disentangle
static and dynamic regions, resulting in poorly estimated deformation fields.
To counteract this, we introduce tuneable dynamic MLPs [36] (i.e. MLPs
with M sets of weights) governed by M sets of learnable blending parameters
{α ∈ RM ×Ng }. The tuning parameters weigh the respective parameters of the
MLP for each input Gaussian i ∈ Ng , enabling the learning of M modes of
variation corresponding to different motion modes. With M = 2, the MLP has
the ability to decouple static and dynamic Gaussians in a smoothly weighted
manner. Given the set of blending parameters {αi }M m=1 for the i-th input Gaus-
sian xi , the output of a single layer of the dynamic MLP y i can be written as a
weighted sum over the M sets of weights:

\boldsymbol {y}_i = \phi \left ( \sum ^M_{m=1} \left ( \alpha _{i,m} \boldsymbol {w}_m^T \boldsymbol {x}_i + \alpha _{i,m} \boldsymbol {b}_m \right ) \right ), (5)

where w and b are the weights and bias, and ϕ is a non-linear activation function.
The blending parameters {αi }M m=1 linearly blend M sets of weights and biases
of each layer of the MLP for each input Gaussian which, when passed through
the activation function, enables nonlinear interaction between α and the output.
Applying α in this way naturally enables the learning different of motion modes
with a single forward pass of the MLP. Thus the dynamic MLP Fθdyn becomes
a function of position, time, and blending parameters α:

\Delta \boldsymbol {x}(t), \Delta \boldsymbol {r}(t), \Delta \boldsymbol {s}(t) = \mathcal {F}^{dyn}_{\theta } (\gamma ({\boldsymbol {x}}), \gamma (t), \boldsymbol {\alpha }). (6)
We implement the dynamic MLP as a single batch matrix multiplication
(see supplementary). Blending parameters α are initialized as a binary mask of
dynamic Gaussians (0-static, 1-dynamic) as follows. For a sliding window, we
project all 3D Gaussians into each camera view: u = Πj (x), obtaining their 2D-
pixel coordinates. We compute their L1 pixel differences from the central frame
to each frame in the window. If the difference is larger than a threshold, we
label that Gaussian 1, otherwise 0. To be robust to occlusions and mislabelling,
we average the assigned label over all views and frames in the window. If the
average is greater than 0.5, we initialize the Gaussian dynamic, otherwise static.
Once initialized, we set the blending parameter α as a learnable parameter,
and optimize it via back-propagation together with the rest of the system. We
assume α is constant for all frames in a sliding window to reduce the complexity
of the optimization, which is a fair assumption given our adaptive window sam-
pling mechanism. Fig. 3 visualizes the learning of the MLP tuning parameter.
We observe α providing higher weight to Gaussians likely to be dynamic. Fig. 3
(right) plots the magnitudes of resulting displacements ∆x output by the MLP.
8 R. Shaw et al.

Note, α does not have to be entirely accurate, as the MLP learns to adjust
accordingly, yet enables the MLP to handle highly imbalanced scenes (Table 6).

11

0.50.5

00

Fig. 3: Dynamic MLPs with tunable parameters α weigh the parameters of the MLP
for each Gaussian. We show renders from two scenes, left: cook spinach [25] and
right: Train [51]. Shown from left-to-right: image render, tunable α parameters, and
normalized MLP displacements ∆x. Note, α highlights the scene’s dynamic regions.

3.6 Temporal Consistency Fine-tuning

Due to the non-deterministic nature of the 3DGS optimization, independently
trained models for each sliding window may produce slightly different results.
When the resulting renders from each model are played back in sequence, notice-
able flickering can be observed, particularly in novel views (though it’s less no-
ticeable in the training views). Consequently, after training models for all sliding
windows separately, we tackle inter-window temporal flickering by introducing a
fine-tuning step for temporal consistency. This step uses a self-supervising loss
function to aid in smooth transitions between the models of separate windows.
We fine-tune each model for a short period (3000 iterations in our exper-
iments). We initiate the process from the first sliding window and progress
through the sequence sequentially. For a window comprising Nw frames, we load
the trained model and the model from the preceding window, with one over-
lapping frame between them. We then freeze all the parameters of the previous
model. In order to fine-tune the model, as the flickering is mainly observed in
novel views, we randomly sample novel test views in between the training views
by rigidly interpolating the training camera poses Pj = [R|t] ∈ R4×4 in SE(3):

P_{\mathrm {novel}} = \exp _{\mathrm {M}} \sum ^{V}_j \beta _j \log _{\mathrm {M}} ( P_j ) (7)

where expM and logM are the matrix exponential and logarithm [1] respectively,
PV
and βj ∈ [0, 1] is a uniformly sampled weighting such that j βj = 1.
In one fine-tuning step, we render the overlapping frames from a randomly
sampled novel viewpoint using both the model of the current window w and the
w
previous window w −1. This means we use the first frame of current window It=0
w−1
and the last frame of the previous window It=Nw −1 . We then apply a consistency
loss, which is simply the L1 loss on the two image renders from both models:
 SWinGS 9

\mathcal {L}_{\mathrm {consistency}} = | I^w_{t=0} - I^{w-1}_{t=N_w-1} |_1 (8)

When fine-tuning a model, we only allow the canonical (static) representation
to change, while freezing the weights of the time-dependent dynamic MLP. Since
the canonical set of 3D Gaussians is shared for all frames in a sliding window,
we need to be careful not to negatively impact other frames when refining the
overlapping frame. To address this, we use an alternating strategy for refinement,
enforcing temporal consistency on the overlapping frame 75% of the time and
training with the remaining views and frames as usual for the remaining 25%.

3.7 Implementation Details

We implement our method in PyTorch, building upon the codebase and dif-
ferentiable rasterizer of 3DGS [20]. We initialize each model with a point cloud
obtained from COLMAP [53]. Each dynamic 3DGS model is trained for 15K iter-
ations for all sampled windows in a sequence. The initial 2K iterations comprise
a warm-up stage; training with only the central window frame and freezing the
weights of the MLP, allowing the canonical representation to stabilize. We opti-
mize the Gaussians’ positions, rotations, scaling, opacities and SH coefficients.
Afterwards, we unfreeze the MLP and allow the deformation field and tuning
parameters α to optimize. We find this staggered optimization approach leads
to better convergence. The number of Gaussians densifies for 8K iterations, after
which the number of Gaussians is fixed. Inputs to the MLP are normalized by the
mean and standard deviation of the canonical point cloud post-warm-up stage
before frequency encoding, leading to faster and stabler convergence. We train
with Adam optimizer, using different learning rates for each parameter following
the implementation of [20]. The learning rate for the MLP and α parameters
are set to 1e-4, with the MLP learning rate undergoing exponential decay by
factor 1e-2 in 20K iterations. During the initial optimization phase, due to the
independence of each Gaussian model, we train each model in parallel on eight
32Gb Tesla V100 GPUs to speed up training. Afterwards, we perform temporal
fine-tuning of each model sequentially using a single GPU for 3K iterations each.

Table 2: Quantitative results on Technicolor

dataset [51] at full resolution. Best and
second best results are highlighted.
Method PSNR SSIM LPIPS FPS
DyNeRF [25] 31.80 0.911 0.142 0.02
HypeRreel [2] 32.50 0.902 0.113 0.45
Fig. 4: Comparison in performance
Dynamic3DG [35] 27.02 0.832 0.228 86.96
consistency for Ours and Dy-
Ours 33.65 0.934 0.117 23.79 namic3DG in consecutive frames.
10 R. Shaw et al.

Table 1: Quantitative results on the Neural 3D Video dataset [25], averaged over all
scenes. Best and second best highlighted. Our method performs best overall whilst
enabling real-time frame rates. Per-scene breakdown of results are given in Table 3.

Method PSNR SSIM LPIPS FPS

MixVoxels [61] 30.42 0.923 0.124 4.30
K-Planes [14] 30.63 0.922 0.117 0.33
HexPlane [7] 30.00 0.922 0.113 0.24
HyperReel [2] 30.78 0.931 0.101 3.60
NeRFPlayer [56] 30.70 0.931 0.121 0.10
StreamRF [23] 30.23 0.904 0.177 9.40
4DGS [65] 27.61 0.916 0.135 30.00
Ours 31.10 0.940 0.096 71.51

4 Results

We evaluate our method on two real-world multi-view dynamic benchmarks: the

Neural 3D Video dataset [25] and the Technicolor dataset [51].
Neural 3D Video comprises real-world dynamic scenes captured with a time-
synchronized multi-view system at 2028 × 2704 resolution at 30 FPS. Cam-
era parameters are estimated using COLMAP [53]. We compare our method to
K-Planes [14], HexPlane [7], MixVoxels [61], HyperReel [2], NeRFPlayer [56],
StreamRF [23], and 4DGS [65]. We compute quantitative metrics for the cen-
tral test view at half the original resolution (1014 × 1352) for 300 frames. The
average results for each method over the dataset are given in Table 1, while Ta-
ble 3 provides a breakdown of the per-scene performance. The results show that
our method performs best regarding PSNR and SSIM metrics while offering the
fastest rendering performance. Qualitative results are shown in Fig. 5.
Technicolor captures real dynamic scenes from a synchronized 4 × 4 camera ar-
ray at 2048×1088 resolution. Following [2], we evaluate on the second row second
column camera on five scenes: Birthday, Fabien, Painter, Theater and Trains.
We compare to DynNeRF [25], HyperReel [2] and Dynamic 3D Gaussians [35],
with quantitative and qualitative results in Table 4 and Fig. 6 respectively.

4.1 Ablation Studies

This section provides ablations showing the effectiveness of our sliding window
and self-supervised temporal consistency fine-tuning strategies. As we train in-
dependent models for each window, flickering artifacts can occur in the final
renders. To visualize this, Fig. 8 plots absolute image error between renders
of neighbouring frames (overlapping frames outlined in red). Without temporal
consistency, we observe a spike in absolute error on overlapping frames, result-
ing in undesirable flickering. However, after fine-tuning, the error in overlapping
frames is drastically reduced. In Table 5, we compute per-frame image metrics
and estimate a measure of the temporal consistency using SoTA video quality
 SWinGS 11

Table 3: Quantitative results on the Neural 3D Video dataset [25], evaluated for 300
frames at 1352 × 1014 resolution. †As reported in [2], ‡Natively trained in lower reso-
lution, upscaled. Best and second best results highlighted respectively. Our method
performs competitively in all metrics, usually coming in either first or second place.

Scene Cook Spinach Cut Roast Beef Flame Steak

Method PSNR SSIM LPIPS PSNR SSIM LPIPS PSNR SSIM LPIPS
MixVoxels [61] 31.39 0.931 0.113 31.38 0.928 0.111 30.15 0.938 0.108
K-Planes [14] 31.23 0.926 0.114 31.87 0.928 0.114 31.49 0.940 0.102
HexPlane‡ [7] 31.05 0.928 0.114 30.83 0.927 0.115 30.42 0.939 0.104
HyperReel [2] 31.77 0.932 0.090 32.25 0.936 0.086 31.48 0.939 0.083
NeRFPlayer† [56] 30.58 0.929 0.113 29.35 0.908 0.144 31.93 0.950 0.088
StreamRF [23] 30.89 0.914 0.162 30.75 0.917 0.154 31.37 0.923 0.152
Ours 31.96 0.946 0.094 31.84 0.945 0.099 32.18 0.953 0.087
Sear steak Coffee Martini Flame Salmon
MixVoxels [61] 30.85 0.940 0.103 29.25 0.901 0.147 29.50 0.898 0.163
K-Planes [14] 30.28 0.937 0.104 29.30 0.900 0.134 29.58 0.901 0.132
HexPlane‡ [7] 30.00 0.939 0.105 28.45 0.891 0.149 29.23 0.905 0.088
HyperReel [2] 31.88 0.942 0.080 28.65 0.897 0.129 28.26 0.941 0.136
NeRFPlayer† [56] 29.13 0.908 0.138 31.53 0.951 0.085 31.65 0.940 0.098
StreamRF [23] 31.60 0.925 0.147 28.13 0.873 0.219 28.69 0.872 0.229
Ours 32.21 0.950 0.092 29.16 0.921 0.105 29.25 0.925 0.100

Table 4: Quantitative results on the Technicolor dataset [51] evaluated at full resolu-
tion. Best and second best results highlighted respectively.

Scene Birthday Fabien Painter

Method PSNR SSIM LPIPS PSNR SSIM LPIPS PSNR SSIM LPIPS
DyNeRF [25] 29.20 0.909 0.067 32.76 0.909 0.242 35.95 0.930 0.147
HyperReel [2] 30.79 0.922 0.062 32.28 0.860 0.217 35.68 0.926 0.123
Dynamic3DG [35] 27.06 0.859 0.113 26.34 0.834 0.268 25.18 0.758 0.395
Ours 33.44 0.959 0.042 34.43 0.925 0.171 36.76 0.948 0.128
Scene Theater Train Average
DyNeRF [25] 29.53 0.875 0.188 31.58 0.933 0.067 31.80 0.911 0.142
HyperReel [2] 33.67 0.895 0.104 30.10 0.909 0.061 32.50 0.902 0.113
Dynamic3DG [35] 28.05 0.799 0.277 28.46 0.908 0.088 27.02 0.832 0.228
Ours 29.81 0.884 0.201 33.99 0.957 0.043 33.69 0.934 0.117
12 R. Shaw et al.

MixVoxels K-Planes HyperReel Ours Ground truth

Fig. 5: Qualitative results on Neural 3D Video [25]. Scenes top to bottom: i) coffee
martini, ii) cook spinach, iii) cut roasted beef, iv) flame salmon, v) flame steak.

HyperReel Dynamic3DG Ours Ground truth

Fig. 6: Qualitative results on Technicolor [51]. Scenes top to bottom: i) Birthday, ii)
Painter, iii) Train.
 SWinGS 13

assessment metric FAST-VQA [66], where the quality score is in the range [0,1].
The table provides results averaged over all scenes from the Neural 3D Video
dataset [25]. Although we incur a minor penalty in some per-frame performance
metrics (SSIM and LPIPS), we obtain a significantly higher video quality (VQA)
score. This indicates a substantial improvement in temporal consistency and
overall perceptual video quality, resulting in more pleasing renderings.

Table 5: Ablation on temporal fine-tuning. Results averaged over all scenes from [25].
Temporal consistency is measured using t-LPIPS [10] and FAST-VQA [66] (quality
score in the range [0,1]). Per-frame performance remains fairly constant, but temporal
consistency and overall perceptual video quality is significantly improved.

Method PSNR ↑ SSIM ↑ LPIPS ↓ t-LPIPS ↓ VQA ↑

w/o temporal consistency 32.01 0.956 0.085 0.0129 0.666
w/ temporal consistency 32.05 0.949 0.093 0.0102 0.726
Ground truth - - - - 0.763

Table 6: Ablation on sliding window size vs adaptive on Birthday scene [51]. We show
the improvement from the dynamic MLP. Adaptive chooses window sizes that match,
and in some metrics exceed, the best fixed-size window performance, striking balance
between performance, temporal consistency (t-LPIPS) and training time (GPU hrs).

Window Size No. Windows Train time PSNR SSIM LPIPS t-LPIPS
3 24 16.0 33.12 0.956 0.048 0.0076
9 6 4.0 33.38 0.959 0.043 0.0053
17 3 2.0 33.01 0.956 0.045 0.0049
25 2 1.3 32.97 0.956 0.043 0.0048
49 1 0.7 32.73 0.955 0.047 0.0051
Adaptive 5 3.3 33.44 0.959 0.042 0.0051
Adaptive (w/o dyn. MLP) 5 3.3 32.76 0.957 0.045 0.0062

Table 6 and Fig. 7 show the impact of the sliding-window size hyperparameter
and our adaptive sampling strategy. Adaptive sampling automatically chooses
appropriate window sizes that match and sometimes exceed the best fixed-size
window performance, striking a good balance between performance, temporal
consistency (t-LPIPS [10]) and training time. Adaptive performs the best overall
with fewer windows (less storage requirements). We also show the advantage
gained from using a dynamic MLP vs a regular MLP, which improves all metrics.

5 Conclusion
We have presented a method to render novel views of dynamic scenes by ex-
tending the 3DGS framework. Results show that our method produces high-
14 R. Shaw et al.

w3 w17 w25 w49 Adapt. w/o dyn. MLP Adaptive

Fig. 7: Ablation on sliding window size vs adaptive window sampling (with and without
dynamic MLP) on scenes from the Technicolor dataset [51]. The single-window scenario
(w49) has the shortest training time but unsatisfactory visual quality, while adaptive
windows offer the best balance between quality and training overhead.
w/o temporal const.
w/ temporal const.

Fig. 8: Ablation on temporal fine-tuning: we display the absolute error between neigh-
bouring frame renders, with overlapping frames highlighted red. After fine-tuning, the
error is substantially reduced and overall perceptual video quality is improved.

quality renderings, even with complex motions e.g. flames. Key to our approach
is sliding-window processing, which adaptively partitions sequences into man-
ageable chunks. Processing each window separately, allowing the canonical rep-
resentation and deformation field to vary throughout the sequence, enables us to
handle complex topological changes, and reduces the magnitude and variability
of the 3D scene flow. In contrast, other methods that learn a single representation
for whole sequences are impractical for long sequences and degrade in quality
with increasing length. Introducing an MLP for each window learns the deforma-
tion field from each canonical representation to a set of per-frame 3D Gaussians.
Moreover, learnable tuning parameters help disentangle static and dynamic parts
of the scene, which we find essential for imbalanced scenes. Our ablations show
self-supervised temporal consistency fine-tuning reduces temporal flickering and
improves the overall perceptual video quality, with only a minor impact on per-
frame performance metrics. Overall, our method performs strongly compared to
recent SoTA quantitatively and obtains sharper, temporally-consistent results.
 SWinGS 15

References
1. Alexa, M.: Linear Combination of Transformations. ACM Trans. Graph. (2002)
2. Attal, B., Huang, J.B., Richardt, C., Zollhöfer, M., Kopf, J., O’Toole, M., Kim, C.:
HyperReel: High-fidelity 6-DoF Video with Ray-conditioned Sampling. In: CVPR
(2023)
3. Bansal, A., Vo, M., Sheikh, Y., Ramanan, D., Narasimhan, S.: 4D Visualization of
Dynamic Events from Unconstrained Multi-view Videos. In: CVPR (2020)
4. Bansal, A., Zollhöfer, M.: Neural Pixel Composition for 3D-4D View Synthesis
from Multi-views. In: CVPR (2023)
5. Barron, J.T., Mildenhall, B., Tancik, M., Hedman, P., Martin-Brualla, R., Srini-
vasan, P.P.: Mip-NeRF: A Multiscale Representation for Anti-aliasing Neural Ra-
diance Fields. In: ICCV (2021)
6. Bârsan, I.A., Liu, P., Pollefeys, M., Geiger, A.: Robust Dense Mapping for Large-
scale Dynamic Environments. In: International Conference on Robotics and Au-
tomation (2018)
7. Cao, A., CV, J.: HexPlane: A Fast Representation for Dynamic Scenes. In: CVPR
(2023)
8. Catley-Chandar, S., Shaw, R., Slabaugh, G., Pérez-Pellitero, E.: RoGUENeRF:
A Robust Geometry-consistent Universal Enhancer for NeRF. arXiv preprint
arXiv:2403.11909 (2024)
9. Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensoRF: Tensorial Radiance Fields.
In: ECCV (2022)
10. Chu, M., Xie, Y., Mayer, J., Leal-Taix’e, L., Thuerey, N.: Learning Temporal Co-
herence via Self-supervision for GAN-based Video Generation. ACM Trans. Graph.
39, 75:1 – 75:13 (2020)
11. Dhamo, H., Nie, Y., Moreau, A., Song, J., Shaw, R., Zhou, Y., Pérez-Pellitero, E.:
HeadGaS: Real-time Animatable Head Avatars via 3D Gaussian Splatting. arXiv
preprint arXiv:2312.02902 (2023)
12. Du, Y., Zhang, Y., Yu, H.X., Tenenbaum, J.B., Wu, J.: Neural Radiance Flow for
4D View Synthesis and Video Processing. In: ICCV (2021)
13. Fang, J., Yi, T., Wang, X., Xie, L., Zhang, X., Liu, W., Nießner, M., Tian, Q.:
Fast Dynamic Radiance Fields with Time-aware Neural Voxels. In: SIGGRAPH
Asia (2022)
14. Fridovich-Keil, S., Meanti, G., Warburg, F.R., Recht, B., Kanazawa, A.: K-Planes:
Explicit Radiance Fields in Space, Time, and Appearance. In: CVPR (2023)
15. Hedman, P., Srinivasan, P.P., Mildenhall, B., Barron, J.T., Debevec, P.: Baking
Neural Radiance Fields for Real-time View Synthesis. In: ICCV (2021)
16. Hu, L., Zhang, H., Zhang, Y., Zhou, B., Liu, B., Zhang, S., Nie, L.: GaussianAvatar:
Towards Realistic Human Avatar Modeling from a Single Video via Animatable
3D Gaussians. In: CVPR (2024)
17. Huang, Y.H., Sun, Y.T., Yang, Z., Lyu, X., Cao, Y.P., Qi, X.: SC-GS: Sparse-
Controlled Gaussian Splatting for Editable Dynamic Scenes. In: CVPR (2024)
18. Işık, M., Rünz, M., Georgopoulos, M., Khakhulin, T., Starck, J., Agapito, L.,
Nießner, M.: HumanRF: High-fidelity Neural Radiance Fields for Humans in Mo-
tion. ACM Trans. Graph. 42(4), 1–12 (2023)
19. Joo, H., Soo Park, H., Sheikh, Y.: MAP Visibility Estimation for Large-scale Dy-
namic 3D Reconstruction. In: CVPR (2014)
20. Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3D Gaussian Splatting for
Real-time Radiance Field Rendering. ACM Trans. Graph. 42(4) (2023)
16 R. Shaw et al.

21. Kocabas, M., Chang, J.H.R., Gabriel, J., Tuzel, O., Ranjan, A.: HUGS: Human
Gaussian Splatting. In: CVPR (2024)
22. Li, L., Shen, Z., Wang, Z., Shen, L., Tan, P.: D-NeRF: Neural Radiance Fields for
Dynamic Scenes. In: CVPR (2020)
23. Li, L., Shen, Z., Wang, Z., Shen, L., Tan, P.: Streaming Radiance Fields for 3D
Video Synthesis. In: NeurIPS (2022)
24. Li, R., Tanke, J., Vo, M., Zollhöfer, M., Gall, J., Kanazawa, A., Lassner, C.: TAVA:
Template-free Animatable Volumetric Actors. In: ECCV (2022)
25. Li, T., Slavcheva, M., Zollhöfer, M., Green, S., Lassner, C., Kim, C., Schmidt, T.,
Lovegrove, S., Goesele, M., Newcombe, R., Lv, Z.: Neural 3D Video Synthesis from
Multi-view Video. In: CVPR (2022)
26. Li, Z., Chen, Z., Li, Z., Xu, Y.: Spacetime Gaussian Feature Splatting for Real-time
Dynamic View Synthesis. In: CVPR (2024)
27. Li, Z., Zheng, Z., Wang, L., Liu, Y.: Animatable Gaussians: Learning Pose-
dependent Gaussian Maps for High-fidelity Human Avatar Modeling. In: CVPR
(2024)
28. Li, Z., Niklaus, S., Snavely, N., Wang, O.: Neural Scene Flow Fields for Space-time
View Synthesis of Dynamic Scenes. In: CVPR (2021)
29. Li, Z., Wang, Q., Cole, F., Tucker, R., Snavely, N.: DynIBaR: Neural Dynamic
Image-Based Rendering. In: CVPR (2023)
30. Lin, K.E., Xiao, L., Liu, F., Yang, G., Ramamoorthi, R.: Deep 3D Mask Volume
for View Synthesis of Dynamic Scenes. In: ICCV (2021)
31. Lin, Y., Dai, Z., Zhu, S., Yao, Y.: Gaussian-Flow: 4D Reconstruction with Dynamic
3D Gaussian Particle. In: CVPR (2024)
32. Liu, Y.L., Gao, C., Meuleman, A., Tseng, H.Y., Saraf, A., Kim, C., Chuang, Y.Y.,
Kopf, J., Huang, J.B.: Robust Dynamic Radiance Fields. In: CVPR (2023)
33. Lombardi, S., Simon, T., Saragih, J., Schwartz, G., Lehrmann, A., Sheikh, Y.:
Neural Volumes: Learning Dynamic Renderable Volumes from Images. ACM Trans.
Graph. 38(4) (2019)
34. Luiten, J., Fischer, T., Leibe, B.: Track to Reconstruct and Reconstruct to Track.
IEEE Robotics and Automation Letters 5(2), 1803–1810 (2020)
35. Luiten, J., Kopanas, G., Leibe, B., Ramanan, D.: Dynamic 3D Gaussians: Tracking
by Persistent Dynamic View Synthesis. In: 3DV (2024)
36. Maggioni, M., Tanay, T., Babiloni, F., McDonagh, S.G., Leonardis, A.: Tunable
Convolutions with Parametric Multi-loss Optimization. In: CVPR (2023)
37. Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng,
R.: NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In:
ECCV (2020)
38. Moreau, A., Song, J., Dhamo, H., Shaw, R., Zhou, Y., Pérez-Pellitero, E.: Human
Gaussian Splatting: Real-time Rendering of Animatable Avatars. In: CVPR (2024)
39. Müller, T., Evans, A., Schied, C., Keller, A.: Instant Neural Graphics Primitives
with a Multiresolution Hash Encoding. ACM Trans. Graph. 41(4), 102:1–102:15
(2022)
40. Newcombe, R.A., Fox, D., Seitz, S.M.: DynamicFusion: Reconstruction and Track-
ing of Non-rigid Scenes in Real-time. In: CVPR (2015)
41. Niemeyer, M., Barron, J.T., Mildenhall, B., Sajjadi, M.S.M., Geiger, A., Radwan,
N.: RegNeRF: Regularizing Neural Radiance Fields for View Synthesis from Sparse
Inputs. In: CVPR (2021)
42. Pang, H., Zhu, H., Kortylewski, A., Theobalt, C., Habermann, M.: ASH: Animat-
able Gaussian Splats for Efficient and Photoreal Human Rendering. In: CVPR
(2024)
 SWinGS 17

43. Park, K., Sinha, U., Barron, J.T., Bouaziz, S., Goldman, D.B., Seitz, S.M., Martin-
Brualla, R.: Nerfies: Deformable Neural Radiance Fields. ICCV (2021)
44. Park, K., Sinha, U., Hedman, P., Barron, J.T., Bouaziz, S., Goldman, D.B., Martin-
Brualla, R., Seitz, S.M.: HyperNeRF: A Higher-dimensional Representation for
Topologically Varying Neural Radiance Fields. ACM Trans. Graph. 40(6) (2021)
45. Peng, S., Dong, J., Wang, Q., Zhang, S., Shuai, Q., Zhou, X., Bao, H.: Animatable
Neural Radiance Fields for Modeling Dynamic Human Bodies. In: ICCV (2021)
46. Peng, S., Yan, Y., Shuai, Q., Bao, H., Zhou, X.: Representing Volumetric Videos
as Dynamic MLP Maps. In: CVPR (2023)
47. Qian, S., Kirschstein, T., Schoneveld, L., Davoli, D., Giebenhain, S., Nießner,
M.: GaussianAvatars: Photorealistic Head Avatars with Rigged 3D Gaussians. In:
CVPR (2024)
48. Qian, Z., Wang, S., Mihajlovic, M., Geiger, A., Tang, S.: 3DGS-Avatar: Animatable
Avatars via Deformable 3D Gaussian Splatting. In: CVPR (2024)
49. Reiser, C., Peng, S., Liao, Y., Geiger, A.: KiloNeRF: Speeding up Neural Radiance
Fields with Thousands of Tiny MLPs. In: ICCV (2021)
50. Rong, X., Huang, J.B., Saraf, A., Kim, C., Kopf, J.: Boosting View Synthesis with
Residual Transfer. In: CVPR (2022)
51. Sabater, N., Boisson, G., Vandame, B., Kerbiriou, P., Babon, F., et al.: Dataset
and Pipeline for Multi-view Light-Field Video. In: CVPRW (2017)
52. Saito, S., Schwartz, G., Simon, T., Li, J., Nam, G.: Relightable Gaussian Codec
Avatars. In: CVPR (2024)
53. Schönberger, J.L., Frahm, J.M.: Structure-from-Motion Revisited. In: CVPR
(2016)
54. Shao, R., Zheng, Z., Tu, H., Liu, B., Zhang, H., Liu, Y.: Tensor4D: Efficient Neural
4D Decomposition for High-fidelity Dynamic Reconstruction and Rendering. In:
CVPR (2023)
55. Slavcheva, M., Baust, M., Cremers, D., Ilic, S.: KillingFusion: Non-rigid 3D Re-
construction without Correspondences. In: CVPR (2017)
56. Song, L., Chen, A., Li, Z., Chen, Z., Chen, L., Yuan, J., Xu, Y., Geiger, A.: NeRF-
Player: A Streamable Dynamic Scene Representation with Decomposed Neural Ra-
diance Fields. IEEE Transactions on Visualization and Computer Graphics 29(5),
2732–2742 (2023)
57. Sun, J., Jiao, H., Li, G., Zhang, Z., Zhao, L., Xing, W.: 3DGStream: On-the-fly
Training of 3D Gaussians for Efficient Streaming of Photo-realistic Free-viewpoint
Videos. In: CVPR (2024)
58. Tanay, T., Maggioni, M.: Global Latent Neural Rendering. In: CVPR (2024)
59. Teed, Z., Deng, J.: RAFT: Recurrent All-pairs Field Transforms for Optical Flow.
In: ECCV (2020)
60. Turki, H., Zhang, J.Y., Ferroni, F., Ramanan, D.: SUDS: Scalable Urban Dynamic
Scenes. In: CVPR (2023)
61. Wang, F., Tan, S., Li, X., Tian, Z., Liu, H.: Mixed Neural Voxels for Fast Multi-
view Video Synthesis. In: ICCV (2023)
62. Wang, L., Zhang, J., Liu, X., Zhao, F., Zhang, Y., Zhang, Y., Wu, M., Yu, J., Xu,
L.: Fourier PlenOctrees for Dynamic Radiance Field Rendering in Real-time. In:
CVPR (2022)
63. Wang, Q., Wang, Z., Genova, K., Srinivasan, P., Zhou, H., Barron, J.T., Martin-
Brualla, R., Snavely, N., Funkhouser, T.: IBRNet: Learning Multi-view Image-
Based Rendering. In: CVPR (2021)
18 R. Shaw et al.

64. Weng, C.Y., Curless, B., Srinivasan, P.P., Barron, J.T., Kemelmacher-Shlizerman,
I.: HumanNeRF: Free-viewpoint Rendering of Moving People from Monocular
Video. In: CVPR (2022)
65. Wu, G., Yi, T., Fang, J., Xie, L., Zhang, X., Wei, W., Liu, W., Tian, Q., Wang, X.:
4D Gaussian Splatting for Real-time Dynamic Scene Rendering. In: CVPR (2024)
66. Wu, H., Chen, C., Hou, J., Liao, L., Wang, A., Sun, W., Yan, Q., Lin, W.: FAST-
VQA: Efficient End-to-end Video Quality Assessment with Fragment Sampling.
In: ECCV (2022)
67. Xian, W., Huang, J.B., Kopf, J., Kim, C.: Space-time Neural Irradiance Fields for
Free-viewpoint Video. In: CVPR (2021)
68. Xiang, J., Gao, X., Guo, Y., Zhang, J.: FlashAvatar: High-fidelity Head Avatar
with Efficient Gaussian Embedding. In: CVPR (2024)
69. Xing, W., Chen, J.: Temporal-MPI: Enabling Multi-Plane Images for Dynamic
Scene Modelling via Temporal Basis Learning. In: ECCV (2022)
70. Xu, Y., Chen, B., Li, Z., Zhang, H., Wang, L., Zheng, Z., Liu, Y.: Gaussian Head
Avatar: Ultra High-fidelity Head Avatar via Dynamic Gaussians. In: CVPR (2024)
71. Yang, Z., Yang, H., Pan, Z., Zhang, L.: Real-time Photorealistic Dynamic Scene
Representation and Rendering with 4D Gaussian Splatting. In: ICLR (2024)
72. Yang, Z., Gao, X., Zhou, W., Jiao, S., Zhang, Y., Jin, X.: Deformable 3D Gaussians
for High-fidelity Monocular Dynamic Scene Reconstruction. In: CVPR (2024)
73. Yu, A., Li, R., Tancik, M., Li, H., Ng, R., Kanazawa, A.: PlenOctrees for Real-time
Rendering of Neural Radiance Fields. In: ICCV (2021)
74. Zhao, F., Yang, W., Zhang, J., Lin, P., Zhang, Y., Yu, J., Xu, L.: HumanNeRF:
Efficiently Generated Human Radiance Field from Sparse Inputs. In: CVPR (2022)
75. Zhou, K., Li, W., Wang, Y., Hu, T., Jiang, N., Han, X., Lu, J.: NeRFLiX: High-
quality Neural View Synthesis by Learning a Degradation-driven Inter-viewpoint
MiXer. In: CVPR (2023)
76. Zhou, T., Tucker, R., Flynn, J., Fyffe, G., Snavely, N.: Stereo Magnification: Learn-
ing View Synthesis using Multiplane Images. ACM Trans. Graph. 37 (2018)