Paper Overview
Field: Computer Vision (CV) Authors: Noam Issachar, Dani Lischinski, Raanan Fattal Published: 2025-06-11 arXiv: 2506.08252
Summary
Standard continuous-time generative models rely on a single, monolithic architecture that must navigate vastly different signal ranges — from isotropic noise to complex data distributions. Uniformly deploying a large network across the entire generation timeline is inherently inefficient.
This paper proposes Complexity-Balanced Splitting (CBS), a framework for temporal capacity allocation that distributes the generative workload across multiple specialized sub-networks. The approach is grounded in function approximation theory and de Boor's equidistribution principle.
Key Ideas
- Local complexity is estimated via two complementary measures:
- A spatial measure based on flow Dirichlet energy
- A geometric measure based on sampling trajectory acceleration
- Capacity is allocated across the generation timeline in proportion to estimated complexity.
- Evaluated across multiple architectures and datasets
- CBS consistently improves synthesis quality
- Boosts FID on SiT-XL by approximately 35% without increasing per-step inference cost
Results
Abstract (Original)
Standard continuous-time generative models rely on monolithic architectures. We propose Complexity-Balanced Splitting (CBS), a framework for temporal capacity allocation that distributes the generative workload across multiple specialized sub-networks. CBS improves FID by ~35% on SiT-XL without increasing per-step inference cost.
--- *Auto-collected on 2025-06-11*