N4MC: Neural 4D Mesh Compression

Guodong Chen; Huanshuo Dong; Mallesham Dasari

cs.CVFeb 23, 2026

N4MC: Neural 4D Mesh Compression

Authors: Guodong Chen, Huanshuo Dong, Mallesham Dasari

Organizations: Northeastern University · University of Science and Technology of China

Abstract

We present N4MC, the first 4D neural compression framework to efficiently compress time-varying mesh sequences by exploiting their temporal redundancy. Unlike prior neural mesh compression methods that treat each mesh frame independently, N4MC takes inspiration from inter-frame compression in 2D video codecs, and learns motion compensation in long mesh sequences. Specifically, N4MC converts consecutive irregular mesh frames into regular 4D tensors to provide a uniform and compact representation. These tensors are then condensed using an auto-decoder, which captures both spatial and temporal correlations for redundancy removal. To enhance temporal coherence, we introduce a transformer-based interpolation model that predicts intermediate mesh frames conditioned on latent embeddings derived from tracked volume centers, eliminating motion ambiguities. Extensive evaluations show that N4MC outperforms state-of-the-art in rate-distortion performance, while enabling real-time decoding of 4D mesh sequences. The implementation of our method is available at: https://github.com/frozzzen3/N4MC.

Figures & tables

Figure 2: N4MC overview. N4MC enables neural 4D mesh compression with 4 modules: (1) A TSDF-Def generation module (top left); (2) A trained auto-encoder and auto-decoder pair to condense TSDF-Def tensors (bottom left); (3) A volume tracking module and a latent code mapping network to generate interpolation priors (top right); (4) A light-weight transformer model for 3D interpolation (bottom right).

Figure 3: Cross-section illustration of N4MC’s volume tracking on input 4D mesh frames to obtain sets of volume centers (orange), and conversion of keyframe meshes (gray) into TSDF-Def tensors. For better visualization, we show the $(4\times 8\times 8)$ intermediate cross-sections of the TSDF tensors. N4MC leverages the tracked volume centers to extract priors for interpolation, to guide the interpolation for regions such as the hands and basketball.

Figure 4: Qualitative comparison of our N4MC with baselines at a bitrate around 4 Mbps. For image-based SSIM and PSNR evaluation, lighting effects are disabled, and vertex normals are converted to RGB colors before rendering. The ”Dancer” and ”Basketball player” sequences appear green, while ”Mitch” and ”Thomas” show pink due to their different orientations. We show a front-facing perspective on people here for all datasets for fair evaluation. To save space, we show full meshes for ”Dancer” dataset and only ground truth with zoomed-in regions for ”Basketball player”, ”Mitch”, and ”Thomas”. N4MC outperforms all the baselines with very high quality.

		Dancer				Basketball Player				Mitch				Thomas
Method	Time [ms] $\downarrow$	D2-PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	Bitrate	D2-PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	Bitrate	D2-PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	Bitrate	D2-PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	Bitrate
Draco [ 16 ]	2.22	54.308	0.913	24.08	3.986	57.577	0.904	23.755	4.046	63.696	0.948	27.1	4.108	63.677	0.952	28.07	4.161
KLT [ 56 ]	24.40	58.166	0.9490	28.027	4.064	58.511	0.9431	27.792	4.036	62.469	0.9450	28.344	4.091	62.026	0.9518	29.494	3.870
TVMC [ 8 ]	11.29	65.527	0.9388	30.2838	4.237	61.840	0.8795	25.889	4.438	72.807	0.9639	34.37	4.664	71.064	0.9639	34.328	4.195
NeCGS [ 46 ]	37.39	62.065	0.9458	26.405	3.480	61.228	0.9540	29.764	3.656	69.256	0.9647	32.337	4.298	67.157	0.9670	33.051	3.520
N4MC (Ours)	38.63	67.276	0.9712	33.567	4.514	66.107	0.9671	33.188	4.311	73.129	0.9779	37.274	4.430	72.759	0.9791	37.811	4.269

Table 1: Quantitative comparison with N4MC and baselines in terms of decoding time, D2-PSNR, image-based SSIM, and PSNR with a similar bitrate around 4 Mbps. The best results are shown in orange, and the second-best in yellow. Corresponding qualitative (visual) comparisons are shown in Fig. 4 .

Panel of Figure 5 — Figure 5: Objective rate–distortion (RD) performance comparison of image-based SSIM versus bitrate on MPEG sequences. To get the target bitrates, for TVMC [ 8 ] and Draco [ 16 ] , the quantization parameter $qp$ is varied from 4 to 14. For NeCGS [ 46 ] and our N4MC, TSDF-Def resolutions of 64, 128, and 256 are used. For KLT [ 56 ] , 16, 32, and 128 basis vectors are employed.

Resolution ( $k$ )	D2-PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	Bitrate	Time [ms] $\downarrow$
64	54.986	0.9296	24.721	2.024	4.097
128	59.633	0.9511	29.278	2.308	7.371
256	66.107	0.9671	33.188	4.311	38.537

Table 2: Ablation study at different resolutions, evaluated in terms of D2-PNSR, image-based SSIM, PSNR, and decoding time.

$n$	$N$	D2-PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	Bitrate	Time [ms] $\downarrow$
5	100	68.225	0.9622	31.847	2.442	6.935
10	100	67.632	0.9610	31.494	2.071	6.417
15	100	67.040	0.9597	31.129	1.968	6.227
5	200	67.635	0.9608	31.480	2.283	7.112
5	300	64.484	0.9517	29.521	1.842	6.868

Table 3: Ablation on group size and total frame number .

Panel of Figure 6 — Figure 6: Qualitative comparisons of N4MC with different TSDF-Def resolutions of 64, 128, and 256.

Method	Time [ms] $\downarrow$	D2-PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	Bitrate
Draco [ 16 ]	9.10	57.597	0.899	27.437	14.487
KLT [ 56 ]	24.7	61.450	0.9424	30.479	10.737
TVMC [ 8 ]	36.457	60.385	0.8820	27.366	9.847
NeCGS [ 46 ]	6.89	61.739	0.9440	30.555	4.931
N4MC (Ours)	7.33	61.121	0.9399	29.898	2.469

Table 4: Quantitative comparison on the Mixed dataset at similar quality levels to illustrate bitrate differences.

	Meta Quest 3	Pixel 9	Xiaomi 12S Pro
Overall time [ms]	346.02	144.34	156.16
Model Inference [ms]	235.29	96.28	100.95
Marching Cubes [ms]	110.73	48.06	55.21

Table 5: Decoding time comparison across mobile devices.

Panel of Figure 7 — Figure 7: Objective rate-distortion (RD) performance comparison of geometry-based D2-PSNR versus bitrate on four 4D mesh sequences: ”Dancer”, ”Basketball player”, ”Mitch”, and ”Thomas”. To get the target bitrates, we use the same setting as in Fig. 5 .

Panel of Figure 8 — Figure 8: Objective rate-distortion (RD) performance comparison of image-based PSNR versus bitrate on four 4D mesh sequences: ”Dancer”, ”Basketball player”, ”Mitch”, and ”Thomas”. To get the target bitrates, we use the same setting as in Fig. 5 .

Figure 9: Qualitative comparison of our N4MC with baselines on ”Mixed” dataset.

Figure 10: Qualitative comparison of our N4MC with baselines on a real captured dataset.

Method	$d_{\text{pm}}(\times 10^{-4})\downarrow$	$d_{\text{normal}}\downarrow$	CR $\uparrow$
NPM [ 11 ]	4.12	7.19 $\degree$	61.39
N4MC (ours)	10.42	9.26 $\degree$	89.56

Table 6: Comparisons of N4MC and Neural Progressive Meshes [ 11 ] on Thingi10K.

d_{\text{pm}}

represents the mean point-to-mesh distance.

d_{\text{normal}}

represents the average normal error.

Figure 11: Ablation on volume center latent mapping. Without latent mapping, N4MC could fail.

Figure 12: Complete qualitative comparison of our N4MC with baselines at a bitrate around 4 Mbps.

Figure 13: Qualitative results of our N4MC on synthetic dataset Thingi10K [ 69 ] .

Figure 14: Decoded meshes in Meta Quest 3.

Fast 4D Mesh Generation by Spatio-Temporal Attention Chains
May 19, 2026Dvir Samuel, Yuval Atzmon, Gal Chechik +13D Mesh Generation 4D Reconstruction
NVIDIA Research, Tel-Aviv, Israel · Bar-Ilan University, Ramat-Gan, Israel
Feed-forward Motion In-betweening for Any 4D
Jun 20, 2026Hiroki Nishizawa, Hubert P. H. Shum, Yoshihiro Fukuhara +23D Mesh Generation
Waseda University · AIST · Durham University
Helix4D: Complex 4D Mesh Generation
May 25, 2026Jiraphon Yenphraphai, Jianqi Chen, Jian Wang +63D Mesh Generation Positional Encoding
1Snap · 2Purdue University · 3KAUST

May 19, 2026J/K move · Enter open · S save

Fast 4D Mesh Generation by Spatio-Temporal Attention Chains

Dvir Samuel, Yuval Atzmon, Gal Chechik, Yoni Kasten

NVIDIA Research, Tel-Aviv, Israel · Bar-Ilan University, Ramat-Gan, Israel

3D Mesh Generation 4D Reconstruction Dynamic Scene Reconstruction Temporal Attention

Open paper AskPDF arXiv

4D mesh generation has recently emerged as a powerful paradigm for recovering dynamic 3D structure from videos, but existing methods remain slow, computationally expensive, and difficult to scale to longer sequences. We introduce a training-free approach that accelerates 4D mesh generation while improving temporal correspondence quality. Our key observation is that temporal correspondences emerge inside a 4D backbone long before its generated meshes become visually accurate. We exploit this with a general framework we call Spatio-Temporal Attention Chain which propagates information across space and time. Starting from vertices on an anchor mesh, the chain maps vertices to latent tokens. It then follows temporal correspondences in latent space, and recovers frame-specific vertices through latent-to-vertex attention. This design avoids expensive explicit matching while preserving anchor mesh details and thereby improving dynamic mesh geometry and temporal consistency. Compared to state-of-the-art, our method generates a 4D mesh in 9 seconds, achieving a

13\times

speedup while producing higher-quality results. Moreover, our approach scales to videos up to

16\times

longer without degrading mesh quality. Beyond generation, the improved correspondences enable competitive zero-shot performance on two downstream tasks: 2D object tracking and 4D tracking. We further show that our framework enables reliable camera estimation, a capability not supported by prior 4D mesh generation methods.

Abstract

Explore similar work

Fast 4D Mesh Generation by Spatio-Temporal Attention Chains

Feed-forward Motion In-betweening for Any 4D

Helix4D: Complex 4D Mesh Generation