Complexity-Balanced Diffusion Splitting

• URL: http://arxiv.org/abs/2606.06477v1
• Date: Thu, 04 Jun 2026 17:57:15 GMT
• Title: Complexity-Balanced Diffusion Splitting
• Abstract summary: Complexity-Balanced Splitting (CBS) is a principled framework for temporal capacity allocation.<n>CBS partitions the diffusion timeline into segments of equal approximation burden.<n>CBS consistently improves quality without increasing per-step inference cost.
• Score: 20.558008394592097
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: 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/.

Related papers

• From Atoms to Entropy: Optimal Noise Allocation for Diffusion Training in the Convex Regime [52.459513910242485]
  We develop a general statistical framework for studying optimal noise-level allocation in diffusion training.<n>We show that the optimized training schedule admits an atomic minimizer, concentrated on finitely many noise levels.<n>We then evaluate the entropic schedule in larger-scale experiments, where full schedule optimization is currently intractable.
  arXiv  Detail & Related papers  (2026-07-10T20:59:06Z)
• Efficient Time Series Clustering from Multiscale Reservoir Dynamics with Granular-Ball Anchoring Graph Optimization [21.33441927042172]
  MSRGC-Net is an efficient time-series clustering framework.<n>It integrates multiscale reservoir computing, granular-ball-based anchoring graph construction, and consensus learning.<n>It consistently outperforms state-of-the-art methods in clustering performance.
  arXiv  Detail & Related papers  (2026-06-10T13:42:35Z)
• Chain-of-Trajectories: Unlocking the Intrinsic Generative Optimality of Diffusion Models via Graph-Theoretic Planning [18.527605555018777]
  Chain-of-Trajectories (CoTj) is a train-free framework enabling System 2 deliberative planning.<n>CoTj allocates computational effort to the most challenging generative phases.<n>This work establishes a new foundation for resource-aware, planning-based diffusion modeling.
  arXiv  Detail & Related papers  (2026-03-16T01:28:51Z)
• Parallel Complex Diffusion for Scalable Time Series Generation [50.01609741902786]
  PaCoDi is a spectral-native architecture that decouples generative modeling in the frequency domain.<n>We show that PaCoDi outperforms existing baselines in both generation quality and inference speed.
  arXiv  Detail & Related papers  (2026-02-10T14:31:53Z)
• Multivariate time series prediction using clustered echo state network [4.4778341776682735]
  Echo state networks (ESNs) offer an efficient alternative to conventional recurrent neural networks.<n>We show that CESNs consistently outperform conventional ESNs in terms of predictive accuracy and robustness to noise.<n>Our algorithm works well across diverse real-world datasets such as the stock market, solar wind, and chaotic Rssler system.
  arXiv  Detail & Related papers  (2025-11-28T17:14:15Z)
• PowerGrow: Feasible Co-Growth of Structures and Dynamics for Power Grid Synthesis [75.14189839277928]
  We present PowerGrow, a co-generative framework that significantly reduces computational overhead while maintaining operational validity.<n> Experiments across benchmark settings show that PowerGrow outperforms prior diffusion models in fidelity and diversity.<n>This demonstrates its ability to generate operationally valid and realistic power grid scenarios.
  arXiv  Detail & Related papers  (2025-08-29T01:47:27Z)
• StreamRL: Scalable, Heterogeneous, and Elastic RL for LLMs with Disaggregated Stream Generation [55.75008325187133]
  Reinforcement learning (RL) has become the core post-training technique for large language models (LLMs)<n>StreamRL is designed with disaggregation from first principles to address two types of performance bottlenecks.<n> Experiments show that StreamRL improves throughput by up to 2.66x compared to existing state-of-the-art systems.
  arXiv  Detail & Related papers  (2025-04-22T14:19:06Z)
• Cross Space and Time: A Spatio-Temporal Unitized Model for Traffic Flow Forecasting [16.782154479264126]
  Predicting backbone-temporal traffic flow presents challenges due to complex interactions between temporal factors. Existing approaches address these dimensions in isolation, neglecting their critical interdependencies. In this paper, we introduce Sanonymous-Temporal Unitized Unitized Cell (ASTUC), a unified framework designed to capture both spatial and temporal dependencies.
  arXiv  Detail & Related papers  (2024-11-14T07:34:31Z)
• Federated Learning over Hierarchical Wireless Networks: Training Latency Minimization via Submodel Partitioning [15.311309249848739]
  Hierarchical independent submodel training (HIST) is a new FL methodology that aims to address these issues in hierarchical cloud-edge-client networks.<n>We demonstrate how HIST can be augmented with over-the-air computation (AirComp) to further enhance the efficiency of the model aggregation over the edge cells.
  arXiv  Detail & Related papers  (2023-10-27T04:42:59Z)
• Complexity Matters: Rethinking the Latent Space for Generative Modeling [65.64763873078114]
  In generative modeling, numerous successful approaches leverage a low-dimensional latent space, e.g., Stable Diffusion. In this study, we aim to shed light on this under-explored topic by rethinking the latent space from the perspective of model complexity.
  arXiv  Detail & Related papers  (2023-07-17T07:12:29Z)
• Vertical Federated Learning over Cloud-RAN: Convergence Analysis and System Optimization [82.12796238714589]
  We propose a novel cloud radio access network (Cloud-RAN) based vertical FL system to enable fast and accurate model aggregation. We characterize the convergence behavior of the vertical FL algorithm considering both uplink and downlink transmissions. We establish a system optimization framework by joint transceiver and fronthaul quantization design, for which successive convex approximation and alternate convex search based system optimization algorithms are developed.
  arXiv  Detail & Related papers  (2023-05-04T09:26:03Z)
• Bayesian Complementary Kernelized Learning for Multidimensional Spatiotemporal Data [11.763229353978321]
  We propose a new statistical framework -- Complementary Complementary Kernelized Learning (BCKL) BCKL offers superior performance in providing accurate posterior mean and high-quality uncertainty estimates.
  arXiv  Detail & Related papers  (2022-08-21T22:38:54Z)
• Scalable Spatiotemporally Varying Coefficient Modelling with Bayesian Kernelized Tensor Regression [17.158289775348063]
  Kernelized tensor Regression (BKTR) can be considered a new and scalable approach to modeling processes with low-rank cotemporal structure. We conduct extensive experiments on both synthetic and real-world data sets, and our results confirm the superior performance and efficiency of BKTR for model estimation and inference.
  arXiv  Detail & Related papers  (2021-08-31T19:22:23Z)
• A Generative Learning Approach for Spatio-temporal Modeling in Connected Vehicular Network [55.852401381113786]
  This paper proposes LaMI (Latency Model Inpainting), a novel framework to generate a comprehensive-temporal quality framework for wireless access latency of connected vehicles. LaMI adopts the idea from image inpainting and synthesizing and can reconstruct the missing latency samples by a two-step procedure. In particular, it first discovers the spatial correlation between samples collected in various regions using a patching-based approach and then feeds the original and highly correlated samples into a Varienational Autocoder (VAE)
  arXiv  Detail & Related papers  (2020-03-16T03:43:59Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.