Skip to content

Fractals made Practical: Denoising Diffusion as Partitioned Iterated Function Systems

Conference: CVPR 2026 arXiv: 2603.13069 Code: None Area: Image Generation Keywords: Diffusion Models, Partitioned Iterated Function Systems (PIFS), Fractal Geometry, Jacobian Analysis, Kaplan-Yorke Dimension

TL;DR

This paper proves that the DDIM deterministic reverse chain is essentially a Partitioned Iterated Function System (PIFS), and derives from this framework three computable geometric quantities requiring no model evaluation. It provides a unified, first-principles explanation for the two-phase denoising dynamics of diffusion models, the effectiveness of self-attention, and four empirical design choices (cosine schedule offset, resolution-dependent logSNR shift, Min-SNR loss weighting, and Align Your Steps sampling).

Background & Motivation

1. State of the Field

Diffusion models generate high-quality images through sequential denoising processes, grounded theoretically in continuous-time SDEs or probability-flow ODEs, with global \(\mathcal{W}_2\) distributional convergence guarantees. However, the continuous perspective treats the learned score network as a black box.

2. Limitations of Prior Work

Existing theory fails to structurally explain two key phenomena: (a) why early steps assemble global spatial context while later steps synthesize local detail; and (b) why self-attention is so effective as a generative primitive. Numerous empirical design choices (cosine offset, Min-SNR weighting, etc.) also lack a unified geometric explanation.

3. Root Cause

The continuous SDE/ODE framework provides elegant distributional convergence guarantees but cannot reveal how the discrete sampling chain assembles image structure at each step — a tension between theoretical elegance and structural interpretability.

4. Paper Goals

  • Provide a structural proof for the two-phase dynamics of the DDIM reverse chain
  • Explain the geometric role of self-attention in diffusion models
  • Derive practical design principles from a unified framework that explain existing empirical techniques

5. Starting Point

The paper reinterprets the DDIM deterministic reverse chain \(\Phi = \Phi_1 \circ \cdots \circ \Phi_T\) as a Partitioned Iterated Function System (PIFS) — a classical mathematical structure from fractal image compression that is naturally suited for handling local self-similarity.

6. Core Idea

The Jacobian of each DDIM operator \(\Phi_t\) decomposes into diagonal blocks (intra-patch) and cross blocks (inter-patch), whose contraction/expansion behavior is fully characterized by three closed-form constants depending only on the noise schedule and patch covariance, enabling analysis of denoising dynamics without running the model.

Method

Overall Architecture

The paper establishes a mathematical mapping from DDIM to PIFS, with the following core pipeline:

  1. Contraction Structure Analysis (§3): Derives contraction conditions for the single-step DDIM operator
  2. Two-Phase Structure Analysis (§4): Computes constants from data statistics and architectural properties to explain the two denoising phases
  3. Attractor Geometry (§5): Computes the fractal dimension of the PIFS attractor via the Lyapunov spectrum
  4. Practical Design Principles (§6): Derives three optimization criteria from the PIFS framework, providing a unified explanation for four empirical designs

Key Designs

Design 1: Dual Contraction Conditions (EC) and (PC)

Function: Establishes two contraction conditions for the single-step DDIM operator \(\Phi_t(x) = \frac{\sqrt{\bar\alpha_{t-1}}}{\sqrt{\bar\alpha_t}} x + b_t \hat\varepsilon_\theta(x,t)\).

Mechanism: The Jacobian \(J_x\Phi_t = \frac{\sqrt{\bar\alpha_{t-1}}}{\sqrt{\bar\alpha_t}} I + b_t J_x\hat\varepsilon_\theta\) contains an expansion term (identity scaling \(>1\)) and a contraction term (score correction with \(b_t < 0\)); contractivity depends on the algebraic properties of the score Jacobian.

  • (EC) Euclidean Contraction: A global condition defining the contraction threshold \(L_t^* = \frac{\sqrt{\bar\alpha_{t-1}/\bar\alpha_t} - 1}{|b_t|}\), depending solely on the noise schedule
  • (PC) Block-Max-Norm Contraction: A patch-level condition that decomposes the Jacobian into diagonal blocks \(\kappa_t^{\mathrm{diag}}\) and cross blocks \(\delta_t^{\mathrm{cross}}\), requiring \(\kappa_t^{\mathrm{diag}} + \delta_t^{\mathrm{cross}} < 1\)

Design Motivation: Natural images exhibit local (not global) self-similarity, necessitating patch-level rather than global contraction guarantees — precisely the advantage of classical PIFS over IFS.

Design 2: Directional Suppression Field and Hierarchical Release

Function: Introduces a directional suppression field \(S_{k,t}(x) = |b_t| \langle v_k^{(1)}, [\nabla_x \Delta_t(x)]_{kk} v_k^{(1)} \rangle\) to quantify the non-Gaussian correction applied by the trained score network to each patch.

Mechanism: Under the Gaussian baseline, the diagonal block spectral norm \(f_t(\lambda_k)\) is \(>1\) (expansive) for all CIFAR-10 patches; however, after training, the network learns a suppression field \(S_{k,t} > 0\) that drives the effective Rayleigh quotient below 1. The key theorem (Stratified Crossover, Thm 22) proves that under the Margin Monotonicity condition (MM), low-variance patches release suppression first and high-variance patches later, producing a strictly variance-ordered hierarchical release.

Design Motivation: Explains why diagonal blocks remain \(\approx 1\) in Regime I rather than expanding as the Gaussian baseline predicts, and how patches sequentially "unlock" detail synthesis in Regime II.

Design 3: Kaplan-Yorke Dimension Formula for the Attractor

Function: Derives the fractal dimension of the PIFS attractor and establishes the discrete Moran equation \(\prod_t f_t(\lambda^{**}) = 1\) to solve for the global expansion threshold \(\lambda^{**}\).

Mechanism: Under the assumptions of Gaussian data and block-diagonal covariance, the Lyapunov spectrum is fully determined by the per-step diagonal expansion function \(f_t(\lambda)\). Diagonal directions with \(\lambda_k > \lambda^{**}\) are expansive, and the KY dimension formula is:

\[d_{\mathrm{KY}} = N^+ + \frac{\sum_{k:\lambda_k > \lambda^{**}} n_k \Lambda(\lambda_k)}{|\Lambda_{k^*}^-|}\]

For non-Gaussian data, the suppression-corrected version satisfies \(d_{\mathrm{KY}}^{\mathrm{eff}} \leq d_{\mathrm{KY}}\), as suppression can only reduce attractor dimensionality.

Design Motivation: Provides model-free predictions of attractor dimension, connecting noise schedule design to the geometric properties of the generative manifold.

Loss & Training

  • Collage Analogy (Thm 12): The DSM training objective is equivalent to minimizing PIFS collage error (up to SNR weighting)
  • \(L^2\)–\(\mathcal{W}_1\) Bridge (Thm 14): The training loss controls the Wasserstein-1 distance to the PIFS fixed point
  • PIFS Regularizer (Thm 15): \(\mathcal{L}_{\mathrm{PIFS}}(\theta) = \mathcal{L}(\theta) + \mu_{\mathrm{reg}} \sum_{t,k,j\neq k} \|[J_x\hat\varepsilon_\theta]_{kj}\|_F^2\), which directly enforces the (PC) condition and can be computed efficiently via JVP/VJP

Key Experimental Results

Main Results: Block-Jacobian Decomposition Validates Two-Phase Structure

Theoretical predictions are validated on a pretrained DDPM CIFAR-10 model using 8×8 patches (\(M=16\), \(n_k=192\)) and a 50-step DDIM sampler.

Training step \(t\) \(\hat\kappa_t^{\mathrm{diag}}\) \(\hat\delta_t^{\mathrm{cross}}\) Global \(\hat\kappa_t\) Phase
980 1.0004 0.0007 1.0011 High noise
800 1.0002 0.0008 1.0010 High noise
600 1.0000 0.0853 1.0853 Regime I
400 1.0026 0.1273 1.1300 Regime I
220 1.0325 0.0768 1.1092 Regime II
20 1.2111 0.1858 1.3969 Detail

Key Finding: In Regime I, global expansion is driven entirely by cross-patch coupling (diagonal blocks \(\approx 1\)); in Regime II, diagonal blocks begin expanding and attention localizes.

Attention Entropy and Cross-Patch Coupling

\(t\) Attention Entropy \(H(A_t)\) (nats) \(\hat\delta_t^{\mathrm{cross}}\) Phase
980 4.963 0.00946 High noise
560 4.662 0.09463 Regime I
160 4.541 0.42899 Regime II
20 4.063 2.06175 Detail

\(\hat\delta_t^{\mathrm{cross}}\) increases 218-fold from \(t=980\) to \(t=20\). Spearman \(\rho(H, \hat\delta^{\mathrm{cross}}) = -1.000\) (\(p < 0.001\)), indicating a perfect inverse correlation between coupling and entropy.

Ablation Study

(PC) Condition Crossover Validation

\(t\) Phase Mean margin slack Violation rate
700 Regime I \(-0.003942\) 16/16
200 I/II transition \(-0.000304\) 14/16
160 Regime II \(+0.001382\) 0/16
40 Regime II \(+0.006412\) 0/16

The crossover occurs at \(t \in [160, 200]\), within approximately a 40-step window. (PC) is universally violated in Regime I and universally satisfied in Regime II.

Spearman Correlation for Hierarchical Release

In the crossover interval (\(t=240, 260\)), \(\rho(\hat\lambda_k, \hat\kappa_t^{\mathrm{diag}})\) is negative and significant (\(p \leq 0.047\)), confirming that low-variance patches are released first. Deep in Regime II (\(t=40\)), \(\rho\) returns to positive at \(0.771\) (\(p=0.001\)), recovering the Gaussian ordering.

Suppression-Corrected KY Dimension (CelebA-HQ Experiment)

On the google/ddpm-celebahq-256 model, with \(\lambda_k \in [38.7, 231.7]\), the Gaussian baseline predicts \(d_{\mathrm{KY}} = 12288\) (full-dimensional expansion). However, the suppression-corrected Moran threshold \(\lambda^{***} = 500 \gg \lambda_{\max}\) yields \(d_{\mathrm{KY}}^{\mathrm{eff}} = 0\). Predicted Lyapunov exponent signs match measured signs for all 16 patches (100% agreement).

Key Findings

  1. Score Deviation Scaling: In the high-noise regime, \(\|\Delta_t\|_2 = O(\sqrt{\bar\alpha_t})\); OLS-fitted slope is \(0.95\) (95% CI \([0.88, 1.02]\)), consistent with theoretical predictions
  2. Information Gain–KY Dimension Proportionality: \(\rho(\mathrm{IG}_t, |\Delta d_t|) \geq 0.9999\), with the ratio CV only 3.4%, nearly perfectly satisfying the Cauchy-Schwarz equality condition
  3. Noise Schedule Comparison: The linear schedule achieves the most uniform \(L_t^*\) (CV 0.341); the cosine schedule yields more uniform information gain (CV 0.867 vs. 1.107)

Highlights & Insights

  1. Deep Mathematical Connection: The paper bridges diffusion models and fractal image compression via the PIFS framework, revealing the elegant correspondence that score matching = collage error minimization
  2. Three Model-Free Constants: \(L_t^*\) (contraction threshold), \(f_t(\lambda)\) (diagonal expansion function), and \(\lambda^{**}\) (global expansion threshold) are fully determined by the schedule and data statistics, forming a universal design language
  3. Unified Theory Explaining Four Empirical Designs: The cosine offset improves the weakest link \(L_1^*\) (4×); Min-SNR balances KY dimension growth; the resolution shift preserves the Moran ratio; AYS concentrates steps where \(L_t^*\) is smallest
  4. Geometric Role of Self-Attention: Query tokens = range blocks; key/value tokens = domain blocks; \(A_{kj}\) = soft domain-range pairing. The hard-attention limit exactly recovers classical PIFS structure, and cross-patch coupling \(\delta_t^{\mathrm{cross}}\) is bounded by the attention weights

Limitations & Future Work

  1. Gaussian Patch Assumption: The core analysis relies on a block-diagonal Gaussian covariance assumption; non-Gaussian cases (e.g., texture-rich data) require more refined suppression field modeling
  2. PIFS Regularizer Not Experimentally Validated: The \(\mathcal{L}_{\mathrm{PIFS}}\) regularizer is proposed but not evaluated in end-to-end training experiments; its practical effect remains to be verified
  3. Limited to DDIM Deterministic Sampling: The analysis primarily targets DDIM (probability-flow ODE); applicability to DDPM stochastic sampling is not discussed in detail
  4. Loose Attention Gradient Bound in Regime I: The bound on \(\|\nabla_x A_{k\ell}\|_{\mathrm{op}}\) in Thm 23 is loose in Regime I; a more precise characterization of query/key temperature is left for future work
  5. Effect of Skip Connections: Encoder skip connections in the UNet architecture lie outside the scope of the attention bound; \(\delta_t^{\mathrm{cross,skip}}\) may dominate in Regime II but is not precisely quantified
  • Fractal Image Compression (Jacquin 1992, Barnsley 1988): PIFS and the Collage Theorem form the mathematical foundation of this paper; the authors reinterpret classical coding theory as an analytical tool for generative models
  • Two-Phase Behavior (Raya & Ambrogioni 2023): Prior empirical observations of two-phase phenomena are given a structural proof (contraction vs. expansion) in this work
  • Information-Constant Schedules (Kingma et al. 2021, Chen et al. 2023): This paper proves that IG uniformity is equivalent to KY dimension growth uniformity (Thm 32), endowing the information-theoretic criterion with geometric meaning
  • Align Your Steps (Sabour et al. 2024): Optimizes step allocation via an upper bound on KL divergence; this paper provides a complementary derivation from the contraction margin perspective, and the two are consistent as both are governed by \(\sqrt{v_t}\)

Rating

⭐⭐⭐⭐⭐ A work of exceptional theoretical depth that seamlessly unifies fractal geometry with diffusion models, providing first-principles explanations for multiple previously isolated empirical techniques. The mathematical derivations are rigorous and the experimental validation is comprehensive (CIFAR-10 + CelebA-HQ), offering paradigm-level insights for the understanding and design of diffusion models.