Turbo4DGen is an arXiv paper that claims a 9.7x average speedup for multi-view 4D generation with no quality loss.
The problem: The SCM attention that keeps frames and viewpoints consistent is expensive, causing out-of-memory failures and very long generation times.
The approach: A spatiotemporal cache reuses intermediate attention across denoising steps. It is paired with semantic-aware attention pruning and an adaptive SCM chain bypass scheduler.
The results: The authors report an average 9.7x speedup "without quality degradation" on the ObjaverseDy and Consistent4D datasets. They also call it the first dedicated acceleration framework for 4D generation.
Nearly ten times faster, according to the authors.
Sign in to comment.
TL;DR of the paper, generated automatically.
Turbo4DGen is an arXiv paper that claims a 9.7x average speedup for multi-view 4D generation with no quality loss.
Nearly ten times faster, according to the authors.