Turbo4DGen: Ultra-Fast Acceleration for 4D Generation
cs.GR, cs.AI, cs.CV
Submitted: 2026-01-24
Updated: 2026-10-28
Comments: ICML 2026
Project page: https://noodle-lab.github.io/turbo4dgen
License: http://creativecommons.org/licenses/by/4.0/
The gist: 4D generation, or dynamic 3D content generation, integrates spatial, temporal, and view dimensions to model realistic dynamic scenes…
Terminology
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.
Sources
- Latent-Shift: Latent Diffusion with Temporal Shift for Efficient Text-to-Video Generation
- Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets
- Objaverse: A Universe of Annotated 3D Objects
- SeerAttention: Learning Intrinsic Sparse Attention in Your LLMs
- CogVideo: Large-scale Pretraining for Text-to-Video Generation via Transformers
- Knowledge Diffusion for Distillation
- Consistent4D: Consistent 360{\deg} Dynamic Object Generation from Monocular Video
- Diffusion4D: Fast Spatial-temporal Consistent 4D Generation via Video Diffusion Models
- VideoFusion: Decomposed Diffusion Models for High-Quality Video Generation
- Efficient4D: Fast Dynamic 3D Object Generation from a Single-view Video
- Learning Transferable Visual Models From Natural Language Supervision
- DreamGaussian4D: Generative 4D Gaussian Splatting
- Make-A-Video: Text-to-Video Generation without Text-Video Data
- DreamVideo: High-Fidelity Image-to-Video Generation with Image Retention and Text Guidance
- Attention-Driven Training-Free Efficiency Enhancement of Diffusion Models
- SV4D 2.0: Enhancing Spatio-Temporal Consistency in Multi-View Video Diffusion for High-Quality 4D Generation
Related papers
- AtlasLC: Fast Codec-Ready Compression of Object-Centric 3D Gaussian Splatting
- Velocity-Space 3D Asset Editing
- Physics-Driven Extraction of Dynamic Fluid Asset from Monocular Video with Generative 3DGS
- SCRIPT: Scalable Diffusion Policy with Multi-stage Training for Language-driven Physics-Based Humanoid Control
- Delaunay-Based Nearest Neighbor Search with Grid-Accelerated Queries and Parallel Construction
- Macrofacet Theory for Gaussian Process Statistical Surfaces