denoising step scheduling

The relative difficulty of denoising tasks across timesteps in diffusion models remains controversial. Method: This work systematically quantifies denoising difficulty per timestep, leveraging both the convergence behavior of denoising error and the relative entropy between true and predicted distributions—revealing that early (low-timestep) denoising is significantly more challenging. Building on this insight, we propose a “curriculum learning” paradigm: timesteps are clustered by difficulty and trained progressively in stages, with joint optimization of the noise schedule. Contribution/Results: Our approach departs from conventional parallel full-timestep training, requiring no architectural or loss-function modifications and remaining compatible with diverse diffusion model enhancements. Extensive experiments on unconditional generation, class-conditional generation, and text-to-image synthesis demonstrate substantial improvements in both model performance and convergence speed.

In high-dimensional spaces, suboptimal noise scheduling in diffusion models degrades generative fidelity due to misaligned noise annealing across feature scales. Method: We identify an inherent dichotomy and complementary deficiencies between VP and VE scheduling schemes in reconstructing high- and low-level data features; building upon a stochastic interpolation framework, we propose a task-adaptive dynamic noise scheduling method that enables joint modeling of multi-scale structural information for the first time. Contribution/Results: Theoretical analysis shows that our scheduler reduces the required number of probability flow ODE discretization steps from Θ(√d) to Θ(1). Empirical validation on Gaussian Mixture (GM) and Curie–Weiss (CW) theoretical models demonstrates full recovery of mode distributions and modal asymmetry, with sampling complexity independent of dimensionality—yielding substantial gains in high-dimensional generation efficiency.

This work identifies a systematic “non-denoising” behavior in practical sampling of conditional diffusion models: under text or observational conditioning, their denoising trajectories consistently deviate from the theoretically ideal path, causing inconsistent generation across algorithms such as DDPM and DDIM. To quantify this phenomenon, the authors introduce *Schedule Deviation*, a novel metric that—through empirical analysis and theoretical justification—demonstrates for the first time that this deviation stems from an inherent inductive bias, independent of model capacity or dataset scale. Further, via theoretical analysis and manifold consistency verification, they establish that smoothness priors fundamentally impede alignment across conditional denoising flows. This work provides an interpretable diagnostic tool for conditional diffusion models and advances the development of robust conditional sampling strategies and training paradigms.

This work addresses the limitations of conventional diffusion models, which rely on handcrafted noise schedules that struggle to balance efficiency and generation quality across varying resolutions and often involve redundant steps. The authors propose a spectrum-guided, instance-adaptive noise scheduling method that derives a “tight” noise schedule grounded in the spectral characteristics of images and theoretical bounds of the diffusion process. During inference, this schedule is dynamically applied through conditional sampling. By incorporating spectral information into the selection of noise levels—a first in the field—the approach significantly enhances generation quality at low inference step counts within single-stage, pixel-level diffusion models while reducing unnecessary computation.

Diffusion models are prone to over-memorizing training samples, which compromises the consistency between the generated distribution and the training data. This work addresses this issue by analyzing the denoising process and reveals, for the first time, that biased sampling of timesteps is a key factor driving memorization. To mitigate this, the authors propose a signal-to-noise ratio–guided timestep reweighting strategy that dynamically adjusts the learning focus along the denoising trajectory based on confidence interval widths, thereby enabling an explicit trade-off between memorization and generalization. Experiments on both image and one-dimensional signal generation tasks demonstrate that shifting the learning emphasis toward later denoising steps significantly reduces memorization while improving sample quality and distribution alignment.

This work addresses the limitation of standard autoregressive diffusion language models in constructing an early semantic scaffold during high-noise stages, which hinders planning-sensitive text generation. The authors propose Noise-Dependent Granularity Control (NDGC), a novel approach that directly leverages the noise level as a signal to dynamically modulate generation granularity throughout the denoising process: at high noise levels, coherent multi-word phrases are generated to support coarse-grained semantic planning, while at low noise levels, the model reverts to token-level refinement. Notably, NDGC achieves this coarse-to-fine, planner-like generation without requiring an explicit planner or hierarchical architecture. Experimental results demonstrate that NDGC significantly improves the timeliness of semantic scaffold construction, the coherence of content restoration, and overall generation quality on tasks such as WritingPrompts.

This work addresses the lack of theoretical grounding in noise scheduling for diffusion models, which has hindered the understanding of empirically effective scheduling strategies. The authors formulate the problem for the first time as an optimal control problem, where the Fisher information serves as the state variable and the noise schedule acts as the control input, with the objective of minimizing an upper bound on the KL sampling error. Within this framework, they derive sufficient conditions for achieving near-optimal sampling error and obtain a tunable closed-form expression for the noise schedule. The proposed method unifies and generalizes exponential and sigmoidal schedules, and, after parameter tuning, achieves improved Fréchet Inception Distance (FID) scores on standard image generation benchmarks.

This work systematically investigates the impact of loss weighting strategies and output parameterizations on model performance in flow matching. Through numerical experiments on both synthetic data with controllable geometric structures and real-world images, the study disentangles their interaction effects across varying data manifold dimensions, model architectures, and dataset scales, using PSNR and FID as evaluation metrics. The analysis reveals, for the first time, how the optimal choice of loss weighting and parameterization depends critically on the intrinsic structure of the data. Building on these insights, the authors formulate practical design principles that substantially improve denoising accuracy and generation quality.

This study investigates whether monotonic denoising schedules are necessary in diffusion models. By designing four classes of structured non-monotonic noise schedules, the authors conduct a systematic evaluation across 90 configurations and 42 hyperparameter ablation settings (involving 10–200 function evaluations) on CIFAR-10 using three prominent frameworks: DDPM, EDM, and Flow Matching. They introduce a novel diagnostic metric—the “schedule sensitivity coefficient”—to quantify architectural dependence on schedule monotonicity, revealing that DDPM suffers significant performance degradation, Flow Matching exhibits moderate sensitivity, and EDM remains largely unaffected. The results demonstrate that non-monotonic schedules do not outperform monotonic baselines, thereby affirming the validity and architecture-specific nature of monotonic sampling in diffusion-based generative modeling.