Squeeze3D: Extreme Neural Compression with Latent Space Bridging

Rishit Dagli; Yushi Guan; Sankeerth Durvasula; Mohammadreza Mofayezi; Nandita Vijaykumar

cs.GRJun 9, 2025

Squeeze3D: Extreme Neural Compression with Latent Space Bridging

Authors: Rishit Dagli, Yushi Guan, Sankeerth Durvasula, Mohammadreza Mofayezi, Nandita Vijaykumar

Organizations: University of Toronto

Abstract

We propose Squeeze3D, a novel framework that leverages implicit prior knowledge learnt by existing pre-trained encoders and decoders to compress 3D data at extremely high compression ratios. Our approach bridges the latent spaces between a pre-trained encoder and a pretrained decoder model through trainable mapping networks. Any 3D asset represented as a mesh, point cloud, or radiance field is first encoded by the pre-trained encoder and then transformed (i.e. compressed) into a highly compact latent code by a mapping network. This latent code can effectively be used as an extremely compressed representation of the mesh, point cloud, or radiance field. A mapping network transforms the compressed latent code into the latent space of a powerful generative model; the decoder of this generative model then recreates the original 3D asset (i.e. decompression). Squeeze3D is trained entirely on generated synthetic data and does not require any 3D datasets. The Squeeze3D architecture can be flexibly used with existing pre-trained 3D encoders and existing generative models. It can flexibly support different formats, including meshes, point clouds, and radiance fields. Our experiments demonstrate that Squeeze3D achieves compression ratios of up to 2187 $\times$ for textured meshes, 58.5 $\times$ for point clouds, and more than 650 $\times$ for radiance fields while maintaining visual quality comparable to many existing methods. Squeeze3D only incurs a small compression and decompression latency since it does not involve training object-specific networks to compress an object.

Figures & tables

Figure 2 : Overview of our Method. Squeeze3D bridges arbitrary latent spaces between 3D encoders and generators through trainable mapping networks. During compression, a 3D geometry is encoded and then transformed into a compact representation via the forward mapping network. During decompression, the reverse mapping network converts this representation into the generator’s latent space, which is then used to reconstruct the original geometry.

Symb.	Description	Symb.	Description
$\mathcal{G}$	A 3D geometry in some format	$E$	3D encoder model: $E(\mathcal{G})\mapsto\mathbf{z}_{E}\in\mathbb{R}^{d_{E}}$
$G$	3D decoder model: $G(\mathbf{z}_{G},c)\mapsto\mathcal{G}^{\prime}$	$\mathbf{z}_{E}$	Latent code from the encoder: $\mathbf{z}_{E}\in\mathbb{R}^{d_{E}}$
$\mathbf{z}_{G}$	Latent code for the generator: $\mathbf{z}_{G}\in\mathbb{R}^{d_{G}}$	$\mathbf{z}_{\text{comp}}$	Compressed representation $\mathbf{z}_{\text{comp}}\in\mathbb{R}^{d_{C}}$
$c$	Conditioning information ( e.g. text prompt, image) for $G$	$F_{\theta}^{E}$	Forward mapping network: $F_{\theta}^{E}(\mathbf{z}_{E})\mapsto\mathbf{z}_{\text{comp}}$
$F_{\theta}^{D}$	Reverse mapping network: $F_{\theta}^{D}(\mathbf{z}_{\text{comp}})\mapsto\mathbf{z}_{G}$	$d_{C}$	Dimensionality of compressed representation
$d_{G}$	Dimensionality of generator latent space	$d_{E}$	Dimensionality of encoder latent space

Table 1 : Notation. The notation we use to describe our method.

Figure 3 : Training Squeeze3D . We show an overview of (a) our process of creating synthetic data to train the mapping networks and (b) our process of training the mapping networks.

Figure 4 : Qualitative mesh compression results. We compare Squeeze3D to state-of-the-art methods. Our approach maintains visually important geometric details. Additional qualitative comparisons are provided in Appendix C.4 .

Figure 5 : Qualitative radiance field compression results. We show qualitative results comparing Squeeze3D to state-of-the-art methods. Our approach achieves a significantly higher compression ratio while maintaining visually important geometric details.

Figure 6 : Compression Ratio and Reconstruction Quality Trade-off. We show the trade-off between compression ratio and reconstruction quality (LPIPS) for Squeeze3D and Draco. Each Squeeze3D operating point varies $d_{C}$ and uses a separately trained forward/reverse mapping pair. Squeeze3D maintains high quality even at extreme compression ratios, whereas Draco’s quality degrades significantly as its compression ratio increases.

Figure 7 : Qualitative point-cloud compression. Reconstructions from Squeeze3D and established codecs; quantitative results are reported in Tables 3 and 10 .

Method	CR ( $\times$ ) $\uparrow$	PSNR $\uparrow$	MS-SSIM $\uparrow$	LPIPS $\downarrow$
Shap-E Native Baseline	1.61	25.80	0.9600	0.0410
Squeeze3D (Shap-E)	1607.50	25.92	0.9615	0.0398

Table 5 : Shap-E Comparison. Comparison between Squeeze3D (MeshAnything

\to

Shap-E) and the native Shap-E encoder-decoder.

Method	CR ( $\times$ ) $\uparrow$	PCQM $\uparrow$	PointSSIM $\uparrow$	Chamfer $\downarrow$
LION Native Baseline	3.60	2.1000	0.5100	0.0210
Squeeze3D (LION)	58.50	1.8437	0.4484	0.0285

Table 6 : LION Comparison. Comparison between Squeeze3D (PointNet++

\to

LION) and the native LION encoder-decoder.

Decoder	Total size
Squeeze3D
Shap-E	905 MB
InstantMesh	1.51 GB
OpenLRM	1.04 GB
LION	270 MB
NeRF-MAE	306 MB

Table 8 : Storage size of each decoder used at decompression.

Appendix figures & tables21 assets

Supplementary material from the paper’s appendix.

Appendix

Encoder/Generator Model	Latent Space Size
MeshAnything ( Chen et al., 2025c )	$257\times 1024$
InstantMesh TriPlane Features ( Xu et al., 2024 )	$3\times 80\times 64\times 64$
LION ( Zeng et al., 2022 )	8320
PointNet++ ( Qi et al., 2017 )	1024
NeRF-MAE ( Irshad et al., 2024 )	$256\times 20\times 20\times 20$

Appendix

Table 10 : Latent Space Sizes. Sizes of latent spaces for encoder and generator models we use.

Encoder $\rightarrow$ decoder	$d_{C}$	Latent size (KB)
MeshAnything $\rightarrow$ OpenLRM	1024	4.00
MeshAnything $\rightarrow$ Shap-E	1024	4.00
MeshAnything $\rightarrow$ InstantMesh	770	3.01
PointNet++ $\rightarrow$ LION	512	2.00
PointNet++ $\rightarrow$ LION	1024	4.00
PointNet++ $\rightarrow$ LION	2048	8.00

Appendix

Table 11: Latent Sizes. The sizes of the compressed representations for our experiments.

Decoder	Mapping training (h)	Paired-data creation (h)
InstantMesh	12	20
LION	5	20
NeRF-MAE	15	3

Appendix

Table 12 : One-time Training and Data-creation Costs. These costs are incurred once for the listed encoder-generator setting; subsequent objects use feed-forward inference.

Hyperparameter	LRM	Shap-E	InstantMesh	NeRF-MAE
Training Precision	FP-32	FP-32	FP-32	FP-32
$d_{C}$	1024	1024	770	24000
Dropout	0.35	0.35	0.35	0.2
Epochs	700	700	700	200
Batch Size	16	8	16	4
Optimizer	Muon	Adam	Muon	Muon

Appendix

Table 13 : Training Hyperparameters. We show the training hyperparameters for the mapping networks we train.

Mesh method	Bits/vertex $\downarrow$	Point-cloud method	Bits/point $\downarrow$
Draco ∗ ( Galligan et al., 2018 )	141.06	Draco ‡ ( Galligan et al., 2018 )	29.92
Draco † ( Galligan et al., 2018 )	145.61	Draco § ( Galligan et al., 2018 )	20.88
Draco ‡ ( Galligan et al., 2018 )	147.13	G-PCC ( Schwarz et al., 2018 )	12.60
Draco § ( Galligan et al., 2018 )	157.75	V-PCC ( Liu et al., 2019 )	9.32
Corto ( Lab, 2025 )	21.24	SparsePCGC ( Wang et al., 2023 )	13.89
Neural Subd. ( Liu et al., 2020a )	86.46	Squeeze3D (LION, $d_{C}=512$ )	8.00

Appendix

Table 14: Standardized Rate Measures. Bits per input vertex for mesh compression and bits per input point for point-cloud compression.

Figure 9 : Interpolation. The compressed representations we obtain can also be interpolated. In these examples, we obtain the compressed representation for the leftmost and rightmost meshes and linearly interpolate between them.

Figure 10 : Point-cloud rate–distortion. PointSSIM (x-axis) versus bits/input point (log y-axis) for Squeeze3D and four codecs. Each Squeeze3D point uses separately trained mappings.

Figure 11 : Multi-view visualization of compressed and reconstructed meshes. The consistent appearance across different viewing angles demonstrates that Squeeze3D learns correct transformations between latent spaces and produces coherent 3D reconstructions. This confirms that our compressed representation encodes complete 3D information rather than view-dependent features.

Figure 12 : Squeeze3D preserves geometry details. We show some meshes compressed with Squeeze3D as wireframes. Notice that Squeeze3D preserves many fine-grained geometric details.

Size ( $\mathbf{z}_{\text{comp}}$ )	Compression (ms)	Decompression (ms)	PCQM $\uparrow$	PointSSIM $\uparrow$
512	3.85	12.74	1.4001	0.3640
1024	3.85	12.74	1.4047	0.3640
2048	4.22	13.39	1.3311	0.4249
4096	4.85	14.04	1.8437	0.4318
8192	5.30	14.19	1.5665	0.4473

Appendix

Table 18 : Ablations. Ablation study on compressed representation size. We analyze how varying the dimensionality of the compressed latent space affects compression time, decompression time, and reconstruction quality (measured by PCQM and PointSSIM).

Figure 15 : Failure Cases. We show examples where Squeeze3D fails to accurately reconstruct the input. (Left) Input mesh with highly intricate details or text. (Right) The reconstruction is smoothed, losing the fine-grained text or surface texture.

Case	PSNR ( $\uparrow$ )	MS-SSIM ( $\uparrow$ )	LPIPS ( $\downarrow$ )	Chamfer ( $\downarrow$ )	EMD ( $\downarrow$ )
XYZRGB Dragon	21.13	0.9315	0.0611	0.05154	0.12827
Lucy	19.47	0.8979	0.1043	0.19425	0.35322

Appendix

Table 20: Quantitative failure-case analysis. Image metrics are means over 36 geometry-only views; normalized Chamfer distance and approximate 3D EMD use 2,048 surface samples. Both meshes are independently normalized to a unit box without rigid ICP alignment.

$\sigma$ / bbox diagonal	PSNR ( $\uparrow$ )	MS-SSIM ( $\uparrow$ )	LPIPS ( $\downarrow$ )	Chamfer ( $\downarrow$ )	EMD ( $\downarrow$ )
0 (clean)	27.50	0.9796	0.0274	0.3930	0.0084
0.005	26.49	0.9767	0.0329	0.4517	0.0174
0.010	24.19	0.9663	0.0533	0.6257	0.0204
0.025	20.12	0.9280	0.1369	1.1186	0.0305
0.050	19.07	0.9125	0.1916	1.9091	0.0954
0.100	18.54	0.9046	0.2140	2.0378	0.1107

Appendix

Table 22 : Input-noise sensitivity.

$\sigma$ / input dynamic range	PSNR ( $\uparrow$ )	MS-SSIM ( $\uparrow$ )	LPIPS ( $\downarrow$ )
0 (clean)	26.62	0.9533	0.0743
0.005	26.55	0.9528	0.0752
0.010	26.43	0.9520	0.0768
0.025	26.10	0.9498	0.0805
0.050	25.62	0.9460	0.0869
0.100	24.90	0.9395	0.0980

Appendix

Table 23 : Radiance-field input-noise sensitivity. Gaussian noise scale is expressed relative to the input dynamic range.

Decoder	Shared storage	Original/object	Latent/object	Break-even objects
Shap-E	2.89 GB	6.43 MB	4.00 KB	449
InstantMesh	3.43 GB	6.43 MB	3.01 KB	534
OpenLRM	2.79 GB	6.43 MB	4.00 KB	435
LION ( $d_{C}=512$ )	312.25 MB	117 KB	2.00 KB	2,716
NeRF-MAE	478.92 MB	58.07 MB	93.75 KB	9

Appendix

Table 24: Decompression-side amortization thresholds. Exact break-even calculations including both the frozen decoder and reverse mapping network.

Explore similar work

Smaller and Faster 3DGS via Post-Training Dictionary Learning
May 28, 2026Jiarong Gong, Jonas Unger, Ehsan MiandjiDynamic 3D Gaussian Splatting Large-Scale 3D Editing Dataset
Linköping University Department of Science and technology
CodecSplat: Ultra-Compact Latent Coding for Feed-Forward 3D Gaussian Splatting
May 25, 2026Pengpeng Yu, Runqing Jiang, Qi Zhang +3Feed-Forward 3D Gaussian Splatting Feed-Forward 3D Gaussian
Sun Yat-sen University, China · Pengcheng Laboratory, China · Peking University, China
Towards Practical Lossless Neural Compression for LiDAR Point Clouds
Mar 26, 2026Pengpeng Yu, Haoran Li, Runqing Jiang +4Point Clouds
Sun Yat-sen University, China · Pengcheng Laboratory, China

May 28, 2026, cs.GRJ/K move · Enter open · S save

Smaller and Faster 3DGS via Post-Training Dictionary Learning

Jiarong Gong, Jonas Unger, Ehsan Miandji

Linköping University Department of Science and technology

Dynamic 3D Gaussian Splatting Large-Scale 3D Editing Dataset Post-Training

Open paper AskPDF arXiv

3D Gaussian Splatting (3DGS) is a promising neural scene representation for real-time rendering, but trained models often suffer from large memory footprints, limiting deployment on less powerful devices. Existing compression techniques often lead to architectures with several additional trainable parameters. While achieving outstanding compression ratios, they introduce noticeable drops in image quality. In this work, we introduce the first dictionary-learning-based compression framework for 3DGS. The proposed post-training compression pipeline can be deployed in virtually any 3DGS model without the need for re-training or modifications to existing 3DGS models. Our compression framework is straightforward to implement, yet provides significant compression capabilities, preserves image quality, and improves real-time rendering performance. Across 13 benchmark scenes, our approach achieves an average compression ratio of 3.95x, 3.10x, and 4.55x when applied to 3DGS, 3DGS-MCMC, and PixelGS, respectively. This yields consistent rendering speedups of 23.3%, 24.3%, and 25.3%, while maintaining image quality.

Abstract

Explore similar work

Smaller and Faster 3DGS via Post-Training Dictionary Learning

CodecSplat: Ultra-Compact Latent Coding for Feed-Forward 3D Gaussian Splatting

Towards Practical Lossless Neural Compression for LiDAR Point Clouds