Fractals made Practical: Denoising Diffusion as Partitioned Iterated Function Systems
| Authors | Ann Dooms |
| Year | 2026 |
| Field | Machine Learning |
| arXiv | 2603.13069 |
| Download | |
| Categories | cs.LG, cs.CV, cs.IT |
Abstract
What is a diffusion model actually doing when it turns noise into a photograph? We show that the deterministic DDIM reverse chain operates as a Partitioned Iterated Function System (PIFS) and that this framework serves as a unified design language for denoising diffusion model schedules, architectures, and training objectives. From the PIFS structure we derive three computable geometric quantities: a per-step contraction threshold , a diagonal expansion function and a global expansion threshold λ^{**}. These quantities require no model evaluation and fully characterize the denoising dynamics. They structurally explain the two-regime behavior of diffusion models: global context assembly at high noise via diffuse cross-patch attention and fine-detail synthesis at low noise via patch-by-patch suppression release in strict variance order. Self-attention emerges as the natural primitive for PIFS contraction. The Kaplan-Yorke dimension of the PIFS attractor is determined analytically through a discrete Moran equation on the Lyapunov spectrum. Through the study of the fractal geometry of the PIFS, we derive three optimal design criteria and show that four prominent empirical design choices (the cosine schedule offset, resolution-dependent logSNR shift, Min-SNR loss weighting, and Align Your Steps sampling) each arise as approximate solutions to our explicit geometric optimization problems tuning theory into practice.
Engineering Breakdown
Plain English
This paper provides a geometric framework for understanding how diffusion models work by showing that the DDIM reverse chain (the process that turns noise into images) operates as a Partitioned Iterated Function System (PIFS). The authors derive three computable quantities—a contraction threshold, an expansion function, and a global expansion threshold—that fully characterize how denoising happens without requiring any model evaluation. These quantities mathematically explain why diffusion models behave in two distinct regimes: assembling global context at high noise levels and synthesizing fine details at low noise levels. This unifies the design of diffusion schedules, architectures, and training objectives under one mathematical language.
Core Technical Contribution
The paper's core innovation is identifying that DDIM denoising follows a Partitioned Iterated Function System structure, which is a mathematical framework from dynamical systems theory. This discovery enables deriving three explicit, model-free geometric quantities that completely characterize denoising dynamics—something prior work couldn't compute without running inference. The PIFS framework unifies previously disconnected design choices in diffusion models (scheduling, architecture, loss functions) into a single coherent mathematical language. This is a fundamental shift from treating diffusion as an empirical engineering problem to understanding it as a structured dynamical system with predictable geometric properties.
How It Works
The mechanism begins by recognizing that DDIM's reverse chain (noise-to-image trajectory) can be decomposed into partitions where the denoising function operates as an iterated function system. At each timestep t, the model contracts information toward a learned mean (controlled by the contraction threshold L_t) while also expanding it along certain directions (quantified by the diagonal expansion function f_t(λ)). The expansion function describes how much the model stretches features along principal directions, and the global expansion threshold λ determines when expansion dominates contraction—marking the transition between regimes. In the high-noise regime where expansion dominates, attention naturally becomes diffuse and operates globally across patches to assemble coarse structure. In the low-noise regime where contraction dominates, attention becomes concentrated and refines fine details locally. These quantities can be computed analytically from the learned denoising model without running forward passes.
Production Impact
For production systems, this framework enables principled optimization of diffusion pipelines without expensive trial-and-error scheduling. Engineers can compute the three geometric quantities from a trained model and use them to automatically design better noise schedules and architecture configurations—predicting which models will work well before deployment. This reduces the engineering burden of hyperparameter tuning for new diffusion variants. The model-free computability means you can diagnose denoising behavior and identify bottlenecks without running inference, saving compute costs during architecture exploration. However, the framework requires that your model operates in the DDIM deterministic regime (not stochastic sampling), and computing the geometric quantities still requires a forward pass through the model structure itself, just not through actual image generation. Integration is straightforward since it's a post-hoc analysis tool that doesn't require retraining existing models.
Limitations and When Not to Use This
The framework assumes DDIM's deterministic reverse chain, so it doesn't directly apply to fully stochastic diffusion samplers or other denoising approaches. The paper appears incomplete (abstract cuts off mid-sentence describing the two-regime behavior), leaving open questions about practical guidance—how do engineers actually use these quantities to improve real systems? The derivation of the expansion function f_t(λ) likely requires spectral analysis of the Jacobian, which may be computationally expensive or numerically unstable for large models. The framework characterizes local denoising dynamics but doesn't obviously extend to understanding global properties like diversity, mode coverage, or how well the model handles distribution shift at inference time. Empirical validation is missing from the abstract—we don't know how well predicted geometric quantities correlate with actual generation quality or sample efficiency.
Research Context
This work bridges dynamical systems theory and diffusion models, building on the recent wave of research attempting to provide theoretical foundations for why diffusion works (prior work by Karras et al. and others examined noise schedules and sampling algorithms empirically). It extends understanding beyond previous schedule optimization work by providing an interpretable geometric language rooted in iterated function systems rather than just empirical performance curves. The two-regime behavior explanation formalizes observations that practitioners have made: diffusion models do seem to work in phases, with early steps handling structure and later steps handling detail. This opens research directions into designing custom architectures that better respect the geometric structure at each noise level, and potentially accelerating sampling by exploiting the regime transitions mathematically.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::