Turbo4DGen: Ultra-Fast Acceleration for 4D Generation

Yuanbin Man; Ying Huang; Zhile Ren; Miao Yin

cs.GRJan 24, 2026

Turbo4DGen: Ultra-Fast Acceleration for 4D Generation

Authors: Yuanbin Man, Ying Huang, Zhile Ren, Miao Yin

Organizations: Department of Computer Science Engineering, University of Texas at Arlington, TX, USA · Independent Researcher, CA, USA

Abstract

4D generation, or dynamic 3D content generation, integrates spatial, temporal, and view dimensions to model realistic dynamic scenes, playing a foundational role in advancing world models and physical AI. However, maintaining long-chain consistency across both frames and viewpoints through the unique spatio-camera-motion (SCM) attention mechanism introduces substantial computational and memory overhead, often leading to out-of-memory (OOM) failures and prohibitive generation times. To address these challenges, we propose Turbo4DGen, an ultra-fast acceleration framework for diffusion-based multi-view 4D content generation. Turbo4DGen introduces a spatiotemporal cache mechanism that persistently reuses intermediate attention across denoising steps, combined with dynamically semantic-aware attention pruning and an adaptive SCM chain bypass scheduler, to drastically reduce redundant SCM attention computation. Our experimental results show that Turbo4DGen achieves an average 9.7 $\times$ speedup without quality degradation on the ObjaverseDy and Consistent4D datasets. To the best of our knowledge, Turbo4DGen is the first dedicated acceleration framework for 4D generation.

Figures & tables

Figure 1: Generated examples of Turbo4DGen , in comparison with the baseline, SV4D ( Xie et al., 2025 ) . Our Turbo4DGen completes the above 4D generation examples in only 9.78s and 12.15s , respectively, whereas SV4D requires around 2 mins (110.85s and 118.76s), yielding 11.33 $\times$ and 9.77 $\times$ speedups without sacrificing content quality.

Figure 2: Performance analysis of removing spatial, camera, or motion attention blocks. It is observed that spatial attention block plays a more critical role in the SCM attention chain.

Figure 3: Latency analysis of multiple components in 4D generation ( Xie et al., 2025 ) . The results show that SCM attention is the main bottleneck, accounting for most of the computational overhead.

Figure 4: The outputs of the SCM attention blocks in the final layer (downsampled to 16 $\times$ 16 for visualization) across adjacent denoising steps (sampling every two steps) exhibit a cosine similarity exceeding 95%, indicating strong redundancy between consecutive steps.

Panel of Figure 5 — Figure 5: Visualization of the spatial cross-attention map indicating semantic representations. (a) Two frames from a reference video; (b) The corresponding spatial cross-attention map; (c) Top- $K$ relevant tokens ( dotted in red ) representing semantic features.

Panel of Figure 6 — Figure 6: Visualization of a generated frame example with different refilling methods after token pruning. (a) Baseline; (b) Pruning with zero-value refilling; (c) Pruning with block cache refilling.

Figure 7: Three randomly selected samples from the ObjaverseDy ( Deitke et al., 2022 ) dataset for the baseline. The similarity of the output from the last-#2 chain at steps

t

and

t-1

is computed, along with the SCM average similarity rate (ASR)

V_{\text{ASR}}

(defined in Eq. 12 ) for steps

t-2

and

t-1

(previous). Results highlight sample-dependent dynamic redundancy, with

V_{\text{ASR}}

reflecting the degree of redundancy. (Redline marks the subsequent step where block skipping can begin.)

Figure 8: Overview of Turbo4DGen . During the 4D generation process, Turbo4DGen adopts a multi-level acceleration scheme across denoising steps. Specifically, the attention-level computations for the noise latent $Z^{t-1}$ are skipped by reusing attention outputs from the rolling cache. Then, we identify token-level redundancy to prune the computations for the camera and motion blocks in the next timestamp. Furthermore, as the denoising process proceeds, the entire SCM chain is adaptively bypassed for further acceleration.

Method	LPIPS $\downarrow$	CLIP-S $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	FVD-F $\downarrow$	FVD-V $\downarrow$	FVD-Diag $\downarrow$	FV4D $\downarrow$
Consistent4D ( Jiang et al., 2023 )	0.148	0.899	16.44	0.866	781.38	510.04	782.79	658.31
STAG4D ( Zeng et al., 2024 )	0.155	0.868	16.73	0.867	848.83	539.96	709.52	833.08
DG4D ( Ren et al., 2023 )	0.156	0.874	16.02	0.860	826.72	543.29	761.58	741.99
L4GM ( Ren et al., 2024 )	0.146	0.902	17.65	0.877	805.87	537.46	782.92	666.48
SV4D ( Xie et al., 2025 )	0.122	0.902	18.47	0.884	754.23	436.86	666.59	699.04
Turbo4DGen ( 9.7 $\times$ Speedup )	0.113	0.917	20.27	0.891	626.36	410.93	642.26	549.03

Table 1: Quantitatively experimental results on the ObjaverseDy ( Deitke et al., 2022 ) dataset, in comparison with the baseline, SV4D ( Xie et al., 2025 ) , and other stat-of-the-art methods. The second-best results are underlined .

Method	LPIPS $\downarrow$	CLIP-S $\uparrow$	FVD-F $\downarrow$
Consistent4D ( Jiang et al., 2023 )	0.160	0.87	1133.93
STAG4D ( Zeng et al., 2024 )	0.126	0.91	992.21
4Diffusion ( Zhang et al., 2024 )	0.165	0.88	–
Efficient4D ( Pan et al., 2024 )	0.130	0.92	–
L4GM ( Ren et al., 2024 )	0.120	0.94	691.87
SV4D ( Xie et al., 2025 )	0.118	0.92	732.40

Table 2: Quantitative results on the Consistent4D ( Jiang et al., 2024 ) dataset, in comparison with the baseline, SV4D ( Xie et al., 2025 ) , and other state-of-the-art methods. Note that the results of Turbo4DGen is zero-shot on this dataset.

Figure 9: The latency, speedup, and peak memory evaluation are presented, where “ X ” indicates OOM errors, and

F

and

V

denote the number of generated frames and views, respectively. Methods marked with “*” are not directly applicable to 4D generation, and their code is modified for experimental evaluation.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 10: Illustration of the cache-enhanced semantic-aware pruning mechanism. With the noise latent $\bm{Z}^{t-1}$ as input of the denoising step $t$ , the spatial attention block first generates the attention map representing the semantic importance of tokens, which is then adopted for the pruning of subsequent camera and motion blocks with intensive computation complexity. Moreover, the attention outputs of the camera and motion blocks are refilled from the rolling cache, leading to faster generation speed while preserving semantic-spatiotemporal consistency.

Method	Speedup $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$
SV4D (Baseline)	1.00 $\times$	0.122	18.47
Turbo4DGen w/ pruning	2.28 $\times$	0.106	22.86
Turbo4DGen w/ caching	5.25 $\times$	0.121	19.48
Turbo4DGen w/ bypass	11.80 $\times$	0.149	18.75
Turbo4DGen	9.70 $\times$	0.113	20.27

Appendix

Table 3: Ablation study on different efficiency granularities for acceleration, reporting results in terms of pruning, caching, and bypass. The bold number indicates the best performance.

Warm-up Steps	Speedup $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	CLIP-S $\uparrow$	SSIM $\uparrow$
3	9.70 $\times$	0.113	20.27	0.917	0.891
2	9.75 $\times$	0.115	19.86	0.909	0.885
1	9.81 $\times$	0.118	19.13	0.904	0.883

Appendix

Table 4: Ablation study of warm-up steps, with results reported in efficiency and generation quality. The yellow color indicates our default setting.

Figure 11: Influence of varying pruning rates, with results reported in terms of speedup, memory usage, and PSNR.

Threshold $\alpha$	Speedup $\uparrow$	PSNR $\uparrow$	CLIP-S $\uparrow$	SSIM $\uparrow$
0.80	12.11	15.31	0.838	0.824
0.85	11.73	17.81	0.863	0.855
0.90	9.70	20.27	0.917	0.891
0.95	9.56	20.29	0.919	0.893

Appendix

Table 5: Ablation study on chain bypassing thresholds, with results reported in efficiency and generation quality. The yellow color indicates our default setting.

Method	LPIPS $\downarrow$	CLIP-S $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
Random	0.188	0.563	10.85	0.571
Turbo4DGen	0.130	0.828	17.13	0.849

Appendix

Table 6: Ablation study of different pruning strategies, with results reported in terms of LPIPS, CLIP-S, PSNR, and SSIM.

Figure 12: Visual quality comparison. ( Left ) Monocular input video; ( Middle ) Results of SV4D ( Xie et al., 2025 ) ; ( Right ) Result of Turbo4DGen .

Figure 13: Visual quality comparison. ( Left ) Monocular input video; ( Middle ) Results of SV4D ( Xie et al., 2025 ) ; ( Right ) Result of Turbo4DGen .

4D Scene Generation

4D scene generation creates dynamic 3D worlds with coherent geometry, appearance, and motion across time.

See all 54 papers

Dream4D: Lifting Camera-Controlled I2V towards Spatiotemporally Consistent 4D Generation
Aug 11, 2025Xiaoyan Liu, Kangrui Li, Jiaxin Liu +2Camera-Controlled Video Generation Image-to-Video Generation
The Chinese University of Hong Kong · The Hong Kong Polytechnic University · The University of New South Wales +1
Sculpt4D: Generating 4D Shapes via Sparse-Attention Diffusion Transformers
Apr 23, 2026Minghao Yin, Wenbo Hu, Jiale Xu +2Diffusion Transformer 3D Shape Generation
1The University of Hong Kong · 2ARC Lab, Tencent PCG
Hallo4D: Multi-Modal Hallucination Mitigation for Consistent Spatio-Temporal Generation
Jul 14, 2026Hongbo Wang, Huaibo Huang, Jie Cao +3Multi-View Consistency 3D Scene Generation
MAIS&NLPR, Institute of Automation, Chinese Academy of Science, Beijing, 100190, China · University of Chinese Academy of Sciences, Beijing, 101408, China · Shanghaitech University, Shanghai, 201210, China

Aug 11, 2025J/K move · Enter open · S save

Dream4D: Lifting Camera-Controlled I2V towards Spatiotemporally Consistent 4D Generation

Xiaoyan Liu, Kangrui Li, Jiaxin Liu, Yuehao Song+1

The Chinese University of Hong Kong · The Hong Kong Polytechnic University · The University of New South Wales+1

Camera-Controlled Video Generation Image-to-Video Generation 4D Reconstruction Dynamic Scene Reconstruction Video Generation 4D Scene Generation

Open paper AskPDF arXiv

The synthesis of spatiotemporally coherent 4D content presents fundamental challenges in computer vision, requiring simultaneous modeling of high-fidelity spatial representations and physically plausible temporal dynamics. Current approaches often struggle to maintain view consistency while handling complex scene dynamics, particularly in large-scale environments with multiple interacting elements. This work introduces Dream4D, a novel framework that bridges this gap through a synergy of controllable video generation and neural 4D reconstruction. Our approach seamlessly combines a two-stage architecture: it first predicts optimal camera trajectories from a single image using few-shot learning, then generates geometrically consistent multi-view sequences via a specialized pose-conditioned diffusion process, which are finally converted into a persistent 4D representation. This framework is the first to leverage both rich temporal priors from video diffusion models and geometric awareness of the reconstruction models, which significantly facilitates 4D generation and shows higher quality (e.g., mPSNR, mSSIM) over existing methods.

Abstract

4D Scene Generation

Explore similar work

Dream4D: Lifting Camera-Controlled I2V towards Spatiotemporally Consistent 4D Generation

Sculpt4D: Generating 4D Shapes via Sparse-Attention Diffusion Transformers

Hallo4D: Multi-Modal Hallucination Mitigation for Consistent Spatio-Temporal Generation