Paper deep dive
ADATEX4D: adaptive texture capacity allocation for 4D gaussian splatting
De Jiang, Peiqiang Wang, Kehong Yuan, Shaohua Ma
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 9/26/2026, 2:38:01 AM
Summary
The paper introduces AdaTex4D, an adaptive texture-capacity module for deformation-based 4D Gaussian Splatting. It addresses the inefficiency of uniform texture resolution by dynamically allocating texture capacity based on visibility-normalized screen-space gradients, temporal peak demand, and deformed local scales. Experiments on N3DV and PanopticSports datasets demonstrate that AdaTex4D reduces texture storage by over 50% while preserving or improving reconstruction quality compared to uniform texture assignment and other baselines like 4DGaussians and Spacetime Gaussian Feature Splatting.
Entities (11)
Relation Signals (7)
AdaTex4D → evaluatedon → N3DV
confidence 95% · Experiments on N3DV and PanopticSports show that AdaTex4D reduces texture storage
AdaTex4D → evaluatedon → PanopticSports
confidence 95% · Experiments on N3DV and PanopticSports show that AdaTex4D reduces texture storage
AdaTex4D → reduces → texture storage
confidence 95% · AdaTex4D reduces texture storage by more than half while preserving reconstruction quality.
AdaTex4D → uses → RGBA triplanes
confidence 95% · Each Gaussian carries packed RGBA triplanes whose two axes grow independently according to visibility normalized screen-space gradients and deformed local scales.
AdaTex4D → improves → PSNR
confidence 90% · At 80 MiB, AdaTex4D improves PSNR by 0.175 and 0.205 dB over Uniform Textured 4DGS on N3DV and PanopticSports, respectively.
AdaTex4D → outperforms → Uniform Textured 4DGS
confidence 90% · Under fixed memory budgets, adaptive allocation also improves quality over uniform texture assignment
A2TG → inspired → AdaTex4D
confidence 85% · Static A2TG addresses this mismatch by growing textures from image-space gradients and Gaussian anisotropy... We address these requirements with AdaTex4D.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Textured Gaussians improve local appearance capacity, but assigning the same texture resolution to every primitive wastes storage on low-detail or weakly visible regions. We introduce AdaTex4D, an adaptive texture-capacity module for deformation-based 4D Gaussian Splatting. Each Gaussian carries packed RGBA triplanes whose two axes grow independently according to visibility normalized screen-space gradients and deformed local scales. Experiments on N3DV and PanopticSports show that AdaTex4D reduces texture storage by more than half while preserving reconstruction quality. Under fixed memory budgets, adaptive allocation also improves quality over uniform texture assignment and reduces overall model and peak memory. These results show that dynamic, anisotropic texture allocation provides a more efficient way to distribute local appearance capacity in 4D Gaussian representations.
Tags
Links
- Source: https://arxiv.org/abs/2609.29963v1
- Canonical: https://arxiv.org/abs/2609.29963v1
Trouble viewing inline? Open PDF directly →
Full Text
24,723 characters extracted from source content.
Expand or collapse full text
AdaTex4D: Adaptive Texture Capacity Allocation for 4D Gaussian Splatting De Jiang Peiqiang Wang Kehong Yuan Shaohua Ma †thanks: *Corresponding author: Shaohua Ma (ma.shaohua@sz.tsinghua.edu.cn). Abstract Textured Gaussians improve local appearance capacity, but assigning the same texture resolution to every primitive wastes storage on low-detail or weakly visible regions. We introduce AdaTex4D, an adaptive texture-capacity module for deformation-based 4D Gaussian Splatting. Each Gaussian carries packed RGBA tri-planes whose two axes grow independently according to visibility-normalized screen-space gradients and deformed local scales. Experiments on N3DV and PanopticSports show that AdaTex4D reduces texture storage by more than half while preserving reconstruction quality. Under fixed memory budgets, adaptive allocation also improves quality over uniform texture assignment and reduces overall model and peak memory. These results show that dynamic, anisotropic texture allocation provides a more efficient way to distribute local appearance capacity in 4D Gaussian representations. Index Terms: 4D Gaussian splatting, dynamic view synthesis, textured Gaussians, adaptive representation, storage-efficient representation †address: 1,2,3,4 Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China 1 Introduction Neural radiance fields established high-quality novel-view synthesis through continuous volumetric functions [15], while 3D Gaussian Splatting (3DGS) replaced expensive ray sampling with an explicit differentiable rasterizer [8]. Dynamic Gaussian methods model time-varying scenes through deformation, motion control, or frame-wise updates [20, 21, 12, 4, 18, 13]. These advances improve dynamic geometry and rendering efficiency, but each Gaussian still offers limited spatial variation inside its projected footprint. Per-primitive texturing increases local appearance capacity [1, 17]. This added capacity creates a new allocation problem. Uniform textured Gaussians assign the same square texture to every primitive, although projected footprint, visibility, motion, and local image detail vary substantially. Large textures are therefore also paid for by small, occluded, or visually simple Gaussians, while detail-critical regions are not distinguished from easy regions. Increasing the global texture resolution can improve local detail, but its memory cost grows for every primitive at once. Existing compact representations and compression methods reduce static Gaussian redundancy [14, 10, 16] or dynamic storage [7, 3, 9, 23]. The key question here is where to place texture capacity. Figure 1: Dynamic evidence aggregation in AdaTex4D. Only visible dynamic observations contribute to each Gaussian’s normalized gradient demand and deformed-scale statistics; these aggregated signals determine whether local texture capacity grows along one or both axes. Figure 2: Overview of AdaTex4D. Starting from a frozen 4DGS backbone, we aggregate visibility-normalized image-space gradient demand and deformed Gaussian scales over dynamic observations. These statistics drive anisotropic texture growth along one or both local axes, while each axis remains capped at resolution 4. The resulting heterogeneous RGBA tri-planes are stored in packed tensors and act as bounded residual appearance features, so the geometry, deformation, opacity, SH appearance, and Gaussian count of the backbone remain unchanged. Static A2TG addresses this mismatch by growing textures from image-space gradients and Gaussian anisotropy [2]. Directly transferring the same rule to 4D scenes is insufficient. A dynamic Gaussian can be visible in only part of a sequence, change projected scale after deformation, or become detail-critical for a short interval. Canonical geometry cannot describe these time-varying states, while a simple temporal average can suppress rare but important peaks. Per-frame resizing would have the opposite problem: it reacts to transient noise and repeatedly changes parameter shapes during optimization. Dynamic allocation therefore needs evidence that is both selective and stable. Occluded observations should not contribute as strongly as visible ones; short-lived high-detail events should survive aggregation; and the growth direction should follow deformed, rather than canonical, local scales. At the same time, the geometric backbone and Gaussian count should remain fixed so that any gain can be attributed to appearance-capacity allocation rather than extra primitives. We address these requirements with AdaTex4D. Starting from a trained deformation-based 4DGS model, we freeze geometry, deformation, opacity, spherical-harmonic (SH) appearance, and primitive count, and optimize only a packed residual texture branch. Three RGBA planes start at ×11\!×\!1 and grow anisotropically to at most ×44\!×\!4. Growth is driven by visibility-normalized image-space gradients, temporal peak demand, and deformed local scales aggregated over dynamic observations. This lets texture capacity follow where and when local appearance detail is needed. Across 12 N3DV and PanopticSports scenes, the resulting allocation preserves reconstruction quality with substantially less texture state. Fixed-memory and fixed-Gaussian-count protocols further separate the benefit of adaptive placement from simply increasing texture resolution or primitive count. Packed texel counts describe storage, but checkpoint size, peak GPU memory, and runtime capture different costs. We report these quantities because the current renderer still materializes dense textures, limiting throughput gains from packing. Our contributions are: 1. We extend anisotropic texture allocation to deforming 4D Gaussians using visibility-aware temporal evidence and deformed local scales. 2. At 80 MiB, AdaTex4D improves PSNR by 0.175 and 0.205 dB over Uniform Textured 4DGS on N3DV and PanopticSports, respectively. 3. Across 12 scenes, AdaTex4D uses 47.3% of the uniform texture bytes while preserving reconstruction quality. 2 Method 2.1 Dynamic Gaussian parameterization At time t, the frozen 4DGS backbone provides the center it μ_i^t, rotation itR_i^t, scale its_i^t, opacity αi _i, and SH color iSHc_i SH for Gaussian i. We freeze these quantities, the deformation field, and the Gaussian count, and optimize only the texture branch. A 3D sample x is first mapped to the Gaussian’s local frame, i=diag(it)−1(it)⊤(−it).x_i=diag(s_i^t)^-1(R_i^t)^\! (x- μ_i^t). (1) Each Gaussian carries three RGBA textures ipT_i^p, p∈xy,yz,zxp∈\xy,yz,zx\. We project ix_i onto each plane, map the projected coordinate from [−1,1][-1,1] to that plane’s texture grid, and bilinearly sample it. The three RGB samples are fused into a residual ir_i, while sampled alpha modulates the Gaussian opacity. Rendering is ^=∑iwi[iSH+0.1tanh(i)],wi=α~i∏j<i(1−α~j). I= _iw_i\! [c_i SH+0.1 (r_i) ], w_i= α_i _j<i(1- α_j). (2) All textures start at ×11\!×\!1. Each axis may grow through 1,2,4\1,2,4\, while the uniform control uses ×44\!×\!4 on every plane. 2.2 Visibility-normalized dynamic evidence A2TG uses image-space gradients to allocate texture capacity in static scenes [2]. For visible occurrence k of Gaussian i, we use the projected-center gradient magnitude qi,k=‖∇i,kℒk‖1q_i,k= \| _m_i,kL_k \|_1 (3) as its texture demand. We split visible observations into eight temporal bins and let qibq_i^b denote the mean demand in nonempty bin b. The final allocation score is simply a mixture of persistent and peak demand, gi=12meanb(qib)+12maxbqib.g_i= 12mean_b(q_i^b)+ 12 _bq_i^b. (4) Only visible observations are included. The highest-demand bin also provides the mean deformed scale used by the growth rule below. Signed gradients are unchanged for optimization; the absolute quantity qi,kq_i,k is used only for allocation. Statistics are reduced across workers before resizing. 2.3 Anisotropic texture growth Growth is evaluated at steps 500 and 1000. A Gaussian is resized only when gi>2×10−5g_i>2× 10^-5. For one texture plane, write its current size as (h,w)(h,w) and the corresponding two components of the selected-bin deformed scale as (sh,sw)(s_h,s_w). We then apply the A2TG-style anisotropic rule [2]: (h′,w′)=(2h,w),sh/sw>4,sw<0.01,(h,2w),sw/sh>4,sh<0.01,(2h,2w),otherwise.(h ,w )= cases(2h,w),&s_h/s_w>4,\ s_w<0.01,\\ (h,2w),&s_w/s_h>4,\ s_h<0.01,\\ (2h,2w),&otherwise. cases (5) Each doubled axis is capped at 4. Thus a thin Gaussian grows mainly along one texture axis, while a more isotropic Gaussian grows along both. New texels are initialized by area interpolation, and Adam moments are resized with the same mapping. 2.4 Optimization and storage Texture values are optimized for 5,000 steps at learning rate 0.0025 using ℒ=0.8‖−^‖1+0.2(1−SSIM(,^)).L=0.8\,\|I- I\|_1+0.2 (1-SSIM(I, I) ). (6) Eight data-parallel workers render one view each per step. Packed texture storage scales with the total number of active RGBA texels, plus a small amount of dimension and offset metadata. We report this texture state together with full-model storage, peak allocated memory, training time, and FPS. The current renderer still materializes dense ×44\!×\!4 textures, so storage savings do not yet increase rendering throughput. Figure 3: Qualitative comparison. Each column shows one method. For each scene, the first row is the full frame and the second row is the enlarged ROI. Top: Tennis. Bottom: Sear Steak. 3 Experiments 3.1 Protocol We evaluate six N3DV scenes [11]—coffee martini, cook spinach, cut roasted beef, flame salmon, flame steak, and sear steak—and six PanopticSports scenes from the CMU Panoptic capture system [6]: basketball, boxes, football, juggle, softball, and tennis. N3DV holds out Cam00, whereas PanopticSports follows its official split; all reported images use a white background. Paired runs use seed 6666 and the same precomputed camera schedule. Dataset means are unweighted averages over all six scenes. Metrics are PSNR, LPIPS-Alex [22], structural dissimilarity at data ranges 1.0 and 2.0 for N3DV, SSIM [19] for PanopticSports, packed texture MiB, checkpoint MiB, peak allocated memory, training time, and FPS. We compare 4DGaussians, Uniform Textured 4DGS, and AdaTex4D under two protocols. The fixed-memory study follows A2TG [2] with 80 MiB and 150 MiB total-model budgets. Each method adjusts Gaussian count and appearance capacity to approach the target budget without exceeding it. The fixed-count study uses 25% and 50% of the converged 4DGaussians count with identical Gaussian subsets across methods. Training uses 5,000 steps on both N3DV and PanopticSports with eight views per step. STG [12] is an external projected anchor; PanopticSports uses TC3DGS [5], and fixed-count Mem is count-normalized to active Gaussians. The 12-scene matrix uses one seed; confidence intervals are paired scene-level bootstrap intervals. 3.2 Main quality–storage result Across 12 scenes, AdaTex4D/Textured 4DGS uses 51.7/109.3 MiB texture (ratio 0.473, −52.7%-52.7\%), 178.0/235.6 MiB full-model storage (−24.4%-24.4\%), and 1.16/1.59 GiB peak memory (−27.0%-27.0\%); training is 2.18/1.64 h (+32.9%+32.9\%) and rendering 84.1/100.8 FPS (−16.6%-16.6\%). Quality is similar: PSNR 28.191/28.157 dB (+0.034+0.034 dB; paired 95% bootstrap CI [−0.035,0.113][-0.035,0.113]) and LPIPS differs by +0.0011+0.0011; N3DV/PanopticSports structural metrics are reported separately. The 57.6 MiB decrease in packed texture state is also the dominant source of the checkpoint reduction because the 4DGS backbone is frozen. The large storage change and small metric change indicate that a uniform ×44\!×\!4 texture assigns substantial capacity to primitives that do not need it. AdaTex4D instead concentrates texels on Gaussians that repeatedly produce reconstruction gradients or become important after deformation. The result is therefore a redistribution of appearance capacity rather than a reduction of geometric capacity. 3.3 Comparison under fixed memory budgets We use the same 80 MiB and 150 MiB total-memory budgets for both datasets. Mem denotes total model storage, and each configuration is tuned to approach the target budget from below. Table 1: Quantitative comparison under fixed total-memory budgets. Both datasets use 80 MiB and 150 MiB budgets. Mem is total model storage in MiB; STG is a projected anchor. Colors mark the best three values; Mem is ranked by total storage. N3DV PanopticSports Method PSNR↑ DSSIM1↓ DSSIM2↓ LPIPS↓ Mem↓ PSNR↑ SSIM↑ LPIPS↓ Mem↓ Memory budget = 80 MiB 4DGaussians [20] 26.257 0.0654 0.0380 0.1741 79.8 25.813 0.8590 0.2606 79.9 Uniform Textured 4DGS [1] 26.235 0.0658 0.0383 0.1698 79.9 25.760 0.8568 0.2487 79.8 Spacetime Gaussian Feature Splatting [12] 26.330 0.0639 0.0368 0.1675 79.7 20.008 0.7681 0.4622 79.8 AdaTex4D (Ours) 26.410 0.0649 0.0374 0.1589 79.6 25.965 0.8612 0.2358 79.7 Memory budget = 150 MiB 4DGaussians [20] 26.546 0.0627 0.0357 0.1662 149.9 26.639 0.8844 0.2096 149.9 Uniform Textured 4DGS [1] 27.046 0.0631 0.0362 0.1553 149.9 26.319 0.8729 0.2181 149.8 Spacetime Gaussian Feature Splatting [12] 26.997 0.0585 0.0335 0.1566 149.8 21.234 0.7930 0.4071 149.8 AdaTex4D (Ours) 27.203 0.0627 0.0360 0.1480 149.7 26.532 0.8749 0.2037 149.7 At 80 MiB, AdaTex4D exceeds Uniform Textured 4DGS by 0.175/0.205 dB on N3DV/PanopticSports at nearly identical storage, and exceeds 4DGaussians by 0.153/0.152 dB. On N3DV, STG gives the lowest dssim values, while AdaTex4D gives the highest PSNR and lowest LPIPS; on PanopticSports, AdaTex4D leads all three quality metrics. The gain is strongest in this tight-budget regime because selective growth avoids spending texture capacity on low-demand primitives. At 150 MiB, AdaTex4D reaches 27.203 dB on N3DV and 26.532 dB on PanopticSports, using 149.7 MiB on both datasets. As the budget relaxes, the PanopticSports gap narrows: 4DGaussians has the highest PSNR/SSIM, while AdaTex4D has the lowest LPIPS. The two budgets show that adaptive textures are most useful when memory is the active constraint. 3.4 Comparison at fixed Gaussian counts We next hold the Gaussian subset fixed at 25% or 50% of the converged 4DGaussians count for each scene. Uniform Textured 4DGS and AdaTex4D use the same Gaussian identities as the 4DGaussians control. This protocol isolates the effect of attaching and allocating texture capacity while keeping the primitive subset fixed. Table 2: Quantitative comparison under fixed Gaussian counts. Six-scene means at 25% or 50% of converged 4DGaussians. DSSIM1/2 use ranges 1.0/2.0. Mem is total active-model storage with A2TG-style count normalization. Colors mark the best three values. N3DV PanopticSports Method PSNR↑ DSSIM1↓ DSSIM2↓ LPIPS↓ Mem↓ PSNR↑ SSIM↑ LPIPS↓ Mem↓ Gaussian count = 25% 4DGaussians [20] 21.917 0.1127 0.0762 0.3024 22.2 23.949 0.8211 0.3444 40.0 Uniform Textured 4DGS [1] 23.505 0.1006 0.0647 0.2753 38.4 24.130 0.8103 0.3183 74.9 Spacetime Gaussian Feature Splatting [12] 22.367 0.1057 0.0714 0.2870 49.1 17.199 0.7011 0.6044 106.3 AdaTex4D (Ours) 22.412 0.1131 0.0734 0.3032 31.9 23.980 0.8079 0.3264 60.4 Gaussian count = 50% 4DGaussians [20] 25.073 0.0763 0.0472 0.2066 44.4 25.813 0.8590 0.2606 80.1 Uniform Textured 4DGS [1] 26.085 0.0701 0.0423 0.1864 76.8 25.964 0.8539 0.2354 150.4 Spacetime Gaussian Feature Splatting [12] 25.709 0.0696 0.0430 0.1917 98.1 19.063 0.7390 0.5206 212.7 AdaTex4D (Ours) 25.402 0.0799 0.0489 0.2027 63.5 25.812 0.8525 0.2408 121.5 At 25%/50% Gaussian count, AdaTex4D saves 16.9%/17.3% active-model storage on N3DV and 19.4%/19.2% on PanopticSports versus Uniform Textured 4DGS. Its PSNR gaps are 1.093/0.683 and 0.150/0.152 dB, respectively. The smaller N3DV gap at 50% suggests that texture allocation cannot fully compensate for missing Gaussian support at 25%. Together, the fixed-memory and fixed-count results separate primitive coverage from texture placement: adaptive allocation helps most when Gaussian support is sufficient. 3.5 Qualitative comparison Figure 3 shows AdaTex4D retaining the head outline and racket hoop in Tennis and the steak’s boundary and surface detail in Sear Steak. These features appear blurred in 4DGaussians and Spacetime Gaussian Feature Splatting, but resemble Uniform Textured 4DGS despite AdaTex4D using less texture storage. 3.6 Ablations and allocation behavior Table 3: Ablation at 50% Gaussian count (Mem: model/texture MiB). N3DV Variant PSNR↑ dssim1↓ dssim2↓ LPIPS↓ Mem↓ Ours 27.188 0.0589 0.0326 0.1540 122.4/30.9 w/o Temporal Peak 26.857 0.0608 0.0335 0.1577 123.5/30.9 w/o Anisotropic Allocation 26.727 0.0607 0.0335 0.1571 123.4/32.4 w/o Visibility Normalization 26.543 0.0605 0.0335 0.1596 123.6/31.0 PanopticSports Variant PSNR↑ SSIM↑ LPIPS↓ Mem↓ Ours 26.854 0.8872 0.1820 236.2/70.0 w/o Temporal Peak 26.532 0.8858 0.1840 236.4/70.2 w/o Anisotropic Allocation 26.402 0.8854 0.1869 237.4/70.0 w/o Visibility Normalization 26.202 0.8854 0.1860 238.5/70.2 Removing temporal peaks lowers PSNR by 0.331/0.322 dB on N3DV/PanopticSports; removing anisotropic allocation lowers it by 0.461/0.452 dB. Removing visibility normalization lowers PSNR by 0.645/0.652 dB. Peaks preserve brief detail, anisotropic growth avoids unnecessary texels, and visibility normalization corrects observation-frequency bias. Square growth uses 32.4 rather than 30.9 MiB on N3DV and lowers PSNR. Across 12 scenes, 53.8% of Gaussians use an anisotropic plane; AdaTex4D stores 44.2% of uniform texels and 47.3% of packed texture bytes. Dimension and offset metadata account for the difference. 4 Conclusion Uniform textures waste capacity when dynamic Gaussians need different levels of detail. AdaTex4D grows packed textures anisotropically using visibility-normalized gradients, temporal peaks, and deformed scales while freezing the 4DGS backbone. Experiments show comparable reconstruction quality with less storage, and ablations support the allocation signals. Dense texture materialization currently limits runtime gains. References [1] B. Chao, H. Tseng, L. Porzi, C. Gao, T. Li, Q. Li, A. Saraf, J. Huang, J. Kopf, G. Wetzstein, and C. Kim (2025) Textured gaussians for enhanced 3D scene appearance modeling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 8964–8974. Cited by: §1, Table 1, Table 1, Table 2, Table 2. [2] S. Hsu, T. Yen, S. Hung, and H. Chu (2026) A2^2TG: adaptive anisotropic textured gaussians for efficient 3D scene representation. In International Conference on Learning Representations, Cited by: §1, §2.2, §2.3, §3.1. [3] Q. Hu, Z. Zheng, H. Zhong, S. Fu, L. Song, X. Zhang, G. Zhai, and Y. Wang (2025) 4DGC: rate-aware 4D gaussian compression for efficient streamable free-viewpoint video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 875–885. Cited by: §1. [4] Y. Huang, Y. Sun, Z. Yang, X. Lyu, Y. Cao, and X. Qi (2024) SC-GS: sparse-controlled gaussian splatting for editable dynamic scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 4220–4230. Cited by: §1. [5] S. Javed, A. J. Khan, C. Dumery, C. Zhao, and M. Salzmann (2025) Temporally compressed 3D gaussian splatting for dynamic scenes. In British Machine Vision Conference (BMVC), Cited by: §3.1. [6] H. Joo, T. Simon, X. Li, H. Liu, L. Tan, L. Gui, S. Banerjee, T. Godisart, B. Nabbe, I. Matthews, T. Kanade, S. Nobuhara, and Y. Sheikh (2019) Panoptic studio: a massively multiview system for social interaction capture. IEEE Transactions on Pattern Analysis and Machine Intelligence 41 (1), p. 190–204. External Links: Document Cited by: §3.1. [7] K. Katsumata, D. M. Vo, and H. Nakayama (2024) A compact dynamic 3D gaussian representation for real-time dynamic view synthesis. In European Conference on Computer Vision, p. 394–412. Cited by: §1. [8] B. Kerbl, G. Kopanas, T. Leimkühler, and G. Drettakis (2023) 3D gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics 42 (4), p. 139:1–139:14. External Links: Document Cited by: §1. [9] S. Kwak, J. Kim, J. Y. Jeong, W. Cheong, J. Oh, and M. Kim (2025) MoDec-GS: global-to-local motion decomposition and temporal interval adjustment for compact dynamic 3D gaussian splatting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 11338–11348. Cited by: §1. [10] J. C. Lee, D. Rho, X. Sun, J. H. Ko, and E. Park (2024) Compact 3D gaussian representation for radiance field. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 21719–21728. Cited by: §1. [11] T. Li, M. Slavcheva, M. Zollhoefer, S. Green, C. Lassner, C. Kim, T. Schmidt, S. Lovegrove, M. Goesele, R. Newcombe, and Z. Lv (2022) Neural 3D video synthesis from multi-view video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 5521–5531. Cited by: §3.1. [12] Z. Li, Z. Chen, Z. Li, and Y. Xu (2024) Spacetime gaussian feature splatting for real-time dynamic view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 8508–8520. Cited by: §1, §3.1, Table 1, Table 1, Table 2, Table 2. [13] Y. Lin, Z. Dai, S. Zhu, and Y. Yao (2024) Gaussian-flow: 4D reconstruction with dynamic 3D gaussian particle. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 21136–21145. Cited by: §1. [14] T. Lu, M. Yu, L. Xu, Y. Xiangli, L. Wang, D. Lin, and B. Dai (2024) Scaffold-GS: structured 3D gaussians for view-adaptive rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 20654–20664. Cited by: §1. [15] B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng (2020) NeRF: representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, p. 405–421. Cited by: §1. [16] S. Niedermayr, J. Stumpfegger, and R. Westermann (2024) Compressed 3D gaussian splatting for accelerated novel view synthesis. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 10349–10358. Cited by: §1. [17] V. Rong, J. Chen, S. Bahmani, K. N. Kutulakos, and D. B. Lindell (2025) GStex: per-primitive texturing of 2D gaussian splatting for decoupled appearance and geometry modeling. In Winter Conference on Applications of Computer Vision, p. 3508–3518. Cited by: §1. [18] J. Sun, H. Jiao, G. Li, Z. Zhang, L. Zhao, and W. Xing (2024) 3DGStream: on-the-fly training of 3D gaussians for efficient streaming of photo-realistic free-viewpoint videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 20675–20685. Cited by: §1. [19] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli (2004) Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing 13 (4), p. 600–612. External Links: Document Cited by: §3.1. [20] G. Wu, T. Yi, J. Fang, L. Xie, X. Zhang, W. Wei, W. Liu, Q. Tian, and X. Wang (2024) 4D gaussian splatting for real-time dynamic scene rendering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 20310–20320. Cited by: §1, Table 1, Table 1, Table 2, Table 2. [21] Z. Yang, X. Gao, W. Zhou, S. Jiao, Y. Zhang, and X. Jin (2024) Deformable 3D gaussians for high-fidelity monocular dynamic scene reconstruction. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 20331–20341. Cited by: §1. [22] R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang (2018) The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition, p. 586–595. External Links: Document Cited by: §3.1. [23] X. Zhang, Z. Liu, Y. Zhang, X. Ge, D. He, T. Xu, Y. Wang, Z. Lin, S. Yan, and J. Zhang (2025) MEGA: memory-efficient 4D gaussian splatting for dynamic scenes. In IEEE/CVF International Conference on Computer Vision, p. 27828–27838. Cited by: §1.