Complexity-Balanced Diffusion Splitting

Hugging Face Daily Papers Papers

Summary

Complexity-Balanced Splitting (CBS) partitions the diffusion timeline into segments of equal approximation burden using local complexity measures, improving synthesis quality by ~35% in FID without increasing inference cost.

Standard continuous-time generative models rely on monolithic architectures that must navigate vastly different signal regimes, from isotropic noise to intricate data distributions. While scaling model capacity improves performance, deploying a massive network uniformly across the entire generative timeline is inherently inefficient. In this work, we propose Complexity-Balanced Splitting (CBS), a principled framework for temporal capacity allocation that distributes the generative workload across multiple specialized sub-networks. Grounded in function approximation theory and de Boor's equidistribution principle, CBS partitions the diffusion timeline into segments of equal approximation burden, allocating more representational capacity to regions where the generative dynamics are more difficult to model. To estimate this local complexity, we introduce two complementary and tractable monitor functions: a spatial measure based on the flow's Dirichlet energy, and a geometric measure based on the acceleration of the sampling trajectories. Using a lightweight auxiliary model to estimate these complexity profiles, our approach eliminates the need for heuristic temporal splits or computationally expensive search procedures. Extensive evaluation across multiple architectures (SiT, JiT, and UNet) and datasets demonstrates that CBS consistently improves synthesis quality without increasing per-step inference cost. In particular, CBS improves FID by ~35% on SiT-XL with CFG relative to naive temporal partitioning. Project page is available at https://noamissachar.github.io/CBS/.
Original Article
View Cached Full Text

Cached at: 06/05/26, 06:06 AM

Paper page - Complexity-Balanced Diffusion Splitting

Source: https://huggingface.co/papers/2606.06477

Abstract

Complexity-Balanced Splitting (CBS) allocates generative capacity across specialized sub-networks by partitioning the diffusion timeline based on local complexity measures, improving synthesis quality without increasing inference cost.

Standard continuous-timegenerative modelsrely on monolithic architectures that must navigate vastly different signal regimes, from isotropic noise to intricate data distributions. While scaling model capacity improves performance, deploying a massive network uniformly across the entire generative timeline is inherently inefficient. In this work, we propose Complexity-Balanced Splitting (CBS), a principled framework fortemporal capacity allocationthat distributes the generative workload across multiple specialized sub-networks. Grounded infunction approximation theoryandde Boor’s equidistribution principle, CBS partitions thediffusion timelineinto segments of equal approximation burden, allocating more representational capacity to regions where the generative dynamics are more difficult to model. To estimate this local complexity, we introduce two complementary and tractable monitor functions: a spatial measure based on the flow’sDirichlet energy, and a geometric measure based on the acceleration of thesampling trajectories. Using a lightweight auxiliary model to estimate thesecomplexity profiles, our approach eliminates the need for heuristic temporal splits or computationally expensive search procedures. Extensive evaluation across multiple architectures (SiT, JiT, and UNet) and datasets demonstrates that CBS consistently improves synthesis quality without increasing per-step inference cost. In particular, CBS improvesFIDby ~35% on SiT-XL with CFG relative to naive temporal partitioning. Project page is available at https://noamissachar.github.io/CBS/.

View arXiv pageView PDFProject pageGitHub2Add to collection

Get this paper in your agent:

hf papers read 2606\.06477

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2606.06477 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2606.06477 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2606.06477 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Elucidating the SNR-t Bias of Diffusion Probabilistic Models

Hugging Face Daily Papers

This paper identifies a Signal-to-Noise Ratio timestep (SNR-t) bias in diffusion probabilistic models during inference, where SNR-timestep alignment from training is disrupted at inference time. The authors propose a differential correction method that decomposes samples into frequency components and corrects each separately, improving generation quality across models like IDDPM, ADM, DDIM, EDM, and FLUX with minimal computational overhead.

Dynamic Chunking for Diffusion Language Models

arXiv cs.CL

This paper introduces Dynamic Chunking for Diffusion Language Models (DCDM), which replaces fixed positional blocks in block discrete diffusion with content-defined semantic chunks using a differentiable Chunking Attention mechanism, achieving consistent improvements across scales up to 1.5B parameters.