Flicker-DDPM: Colored Noise Diffusion

• The paper demonstrates that replacing isotropic white noise with flicker colored noise preserves natural image spectral statistics, reducing the need for extensive sampling steps.
• Flicker-DDPM uses a Matérn-based covariance and FFT diagonalization to achieve nearly linear reverse diffusion dynamics in Fourier space, enhancing frequency decoupling.
• Empirical results on CIFAR-10 reveal a significant FID improvement, with a 51.7% reduction at 150 steps versus standard DDPM, evidencing a 3.33× speedup.

Flicker-DDPM is a denoising diffusion model that replaces the isotropic white noise used in standard denoising diffusion probabilistic models (DDPMs) with flicker, or $1/f^\alpha$, colored noise whose spectrum is designed to match the power-law statistics of natural images. The method is motivated by self-organized criticality and by the empirical observation that natural-image power spectra typically follow $P(k)\propto 1/k^\alpha$. In the reported CIFAR-10 experiments, this spectral matching allows Flicker-DDPM to match or surpass a standard DDPM baseline using $3.33$ times fewer sampling steps, while adding negligible computational cost per step (Mao et al., 2 Jun 2026).

1. Formulation of the forward process

Standard DDPMs corrupt data $x_0$ into white Gaussian noise through

$$x_t = \sqrt{\bar\alpha_t}\,x_0 + \sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$

Flicker-DDPM replaces the white-noise covariance $I$ by a covariance $\Sigma$ whose Fourier-domain power spectrum matches the natural-image law $P(k)\propto k^{-\alpha}$. In $\epsilon$-prediction form, the forward process becomes

$$x_t = \sqrt{\bar\alpha_t}\,x_0 + \sqrt{1-\bar\alpha_t}\,L\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I),\qquad \Sigma = L\,L^{\!\top}.$$

The colored noise is generated by imposing a power-law spatial correlation kernel

$P(k)\propto 1/k^\alpha$0

Under the Wiener–Khinchin relation in two dimensions, a covariance $P(k)\propto 1/k^\alpha$1 produces a spectrum $P(k)\propto 1/k^\alpha$2. The exposition reports a naive continuum relation $P(k)\propto 1/k^\alpha$3, but emphasizes that realistic image spectra require using the Matérn covariance with a finite low-$P(k)\propto 1/k^\alpha$4 cutoff. Under that treatment, the relation becomes

$P(k)\propto 1/k^\alpha$5

For CIFAR-10, a log–log regression of the azimuthal spectrum gives $P(k)\propto 1/k^\alpha$6, which yields the leading prediction $P(k)\propto 1/k^\alpha$7. After accounting for finite-grid curvature through an effective midband exponent $P(k)\propto 1/k^\alpha$8, the optimal parameter is predicted as

$P(k)\propto 1/k^\alpha$9

which the paper states is in exact agreement with the value found by direct search (Mao et al., 2 Jun 2026).

2. Spectral matching and the reverse-process theory

The central theoretical claim is that when the reverse diffusion is initialized from noise whose spectrum already matches the data, the denoising dynamics become essentially linear in Fourier space. The paper writes the variance-preserving reverse SDE in Fourier components as

$3.33$0

Because natural-image statistics are treated as approximately translationally invariant, the Jacobian of the learned score function is described as circulant in pixel space and therefore diagonal in Fourier space. The score is then approximated by the ansatz

$3.33$1

where the restoring-force coefficient $3.33$2 is measured by regressing the U-Net output on the noisy input.

Applying Itô’s lemma to the variance of each Fourier mode gives a decoupled linear ODE for the instantaneous power $3.33$3: $3.33$4 Here $3.33$5 denotes the noise spectrum: flat for white noise and proportional to $3.33$6 for colored noise.

Two consequences are highlighted. First, under exact spectral matching, if the forward process ends in noise whose spectrum equals the data, then $3.33$7 for all $3.33$8 and at all $3.33$9, so the reverse process does not need to rebuild the $x_0$0 power law. Second, the paper defines a linearization quality $x_0$1 and reports that white noise leaves low-$x_0$2 dynamics strongly nonlinear, with $x_0$3 at $x_0$4, whereas colored noise yields $x_0$5 uniformly across all modes, consistent with mode-decoupled Ornstein–Uhlenbeck behavior (Mao et al., 2 Jun 2026).

This suggests that the reported acceleration is not attributed merely to a better initialization in pixel space, but to a reduction in frequency-by-frequency spectral reshaping during reverse diffusion.

3. Construction of the colored-noise module

The paper describes two equivalent implementations of the colored-noise module. The first uses Cholesky factorization of the $x_0$6 covariance matrix $x_0$7, yielding $x_0$8 so that $x_0$9. This incurs $$x_t = \sqrt{\bar\alpha_t}\,x_0 + \sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$0 cost once offline.

The second uses FFT-based diagonalization. In this construction, $$x_t = \sqrt{\bar\alpha_t}\,x_0 + \sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$1 is embedded into a circulant matrix, diagonalized by FFT, and sampled through

$$x_t = \sqrt{\bar\alpha_t}\,x_0 + \sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$2

with $$x_t = \sqrt{\bar\alpha_t}\,x_0 + \sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$3 cost per sample.

In practice, the FFT approach is used so that each forward or reverse step requires only one batched FFT multiply, which the paper characterizes as negligible relative to the U-Net cost. White-noise DDPM is recovered as the special case $$x_t = \sqrt{\bar\alpha_t}\,x_0 + \sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$4 (Mao et al., 2 Jun 2026).

A plausible implication is that the method is architecturally minimal: the principal change is the substitution $$x_t = \sqrt{\bar\alpha_t}\,x_0 + \sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$5 in the forward corruption model and the corresponding matched noise process in sampling.

4. Empirical results on CIFAR-10

The reported experiments train a standard U-Net on CIFAR-10 at $$x_t = \sqrt{\bar\alpha_t}\,x_0 + \sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$6 RGB resolution for 200 epochs with a linear $$x_t = \sqrt{\bar\alpha_t}\,x_0 + \sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$7-schedule from $$x_t = \sqrt{\bar\alpha_t}\,x_0 + \sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$8 to $$x_t = \sqrt{\bar\alpha_t}\,x_0 + \sqrt{1-\bar\alpha_t}\,\epsilon,\qquad \epsilon\sim\mathcal N(0,I).$$9. Flicker-DDPM uses $I$0, consistent with the spectral analysis (Mao et al., 2 Jun 2026).

The paper compares FID on 10k samples against the number of sampling steps $I$1:

$I$2	White	Flicker
100	36.17	22.57
150	25.36	12.24
200	18.08	11.57
500	13.02	11.96

The corresponding improvements reported in the exposition are $I$3 at $I$4, $I$5 at $I$6, $I$7 at $I$8, and $I$9 at $\Sigma$0. At $\Sigma$1, Flicker-DDPM attains FID $\Sigma$2, surpassing the white-noise DDPM even at $\Sigma$3, where the baseline reaches FID $\Sigma$4. This is the basis for the reported $\Sigma$5 sampling speedup.

The exposition also states that sample-quality comparisons at $\Sigma$6 show Flicker-generated images to be visibly sharper and more coherent, and that ablation over $\Sigma$7 reveals a clear optimum at approximately $\Sigma$8, in agreement with the Matérn-based prediction. Within the scope of the provided results, the empirical picture is therefore consistent with the reverse-process theory: aligning the injected noise spectrum with the data spectrum reduces the step budget required for comparable or better generation quality.

5. Relation to conventional DDPM assumptions

The defining contrast with standard DDPM is the treatment of the forward corruption process. In the conventional setting described in the paper, the terminal state is isotropic white Gaussian noise. Flicker-DDPM instead uses a nontrivial covariance $\Sigma$9 chosen so that the terminal noise retains the same spectral law as the dataset. This changes the role of the diffusion process: rather than first destroying and later reconstructing the full power-law structure of natural images, the model can preserve the correct spectral envelope throughout the stochastic evolution (Mao et al., 2 Jun 2026).

This framing directly targets what the authors call a spectral-reshaping bottleneck. Under white noise, the reverse dynamics must re-establish low-frequency power across modes whose behavior remains comparatively nonlinear. Under colored noise, the spectral target is already present at the terminal distribution, and the reported $P(k)\propto k^{-\alpha}$0 measurements indicate more linear dynamics across frequencies.

A common misconception would be to interpret the method as introducing a large architectural modification or a fundamentally different training target. The provided exposition instead presents it as a minimal alteration to the noise model, with negligible additional per-step computation and full compatibility with the standard U-Net training pipeline used in the CIFAR-10 experiments.

6. Scope, limitations, and extensions

The paper identifies several limitations. A pure power-law kernel underweights very high-$P(k)\propto k^{-\alpha}$1 texture relative to the true image spectrum, which may deviate from a strict $P(k)\propto k^{-\alpha}$2 law at the finest scales. The authors therefore suggest combining flicker noise at low frequencies with a targeted high-frequency boost to recover fine detail without reintroducing spectral mismatch (Mao et al., 2 Jun 2026).

The analytic relation

$P(k)\propto k^{-\alpha}$3

is stated to be specific to $P(k)\propto k^{-\alpha}$4 and to Matérn asymptotics. Extension to other modalities, including audio, protein structures, and astrophysical time series, would require measuring the relevant spectral exponent and selecting an appropriate covariance model. This suggests that Flicker-DDPM is not a universal closed-form prescription for all data types, but a framework in which the noise process is adapted to the spectral statistics of the dataset.

The method is also described as fully orthogonal to existing accelerators such as DDIM, progressive distillation, and consistency models, and thus as combinable with them for further speedups. That statement does not by itself establish combined empirical gains, but it places Flicker-DDPM within a broader line of diffusion acceleration research: rather than shortening trajectories solely through deterministic samplers or distillation, it alters the terminal noise distribution so that the reverse path is spectrally easier to traverse.

7. Significance within diffusion modeling

Flicker-DDPM formalizes a data-spectrum-matched alternative to white-noise diffusion. Its contribution is not only empirical but also analytic: the colored-noise forward process is tied to a spatial correlation kernel $P(k)\propto k^{-\alpha}$5, the parameter $P(k)\propto k^{-\alpha}$6 is connected to the observed spectral exponent $P(k)\propto k^{-\alpha}$7, and the reverse dynamics are analyzed through a frequency-domain linear theory that predicts improved mode decoupling and reduced nonlinear spectral reconstruction (Mao et al., 2 Jun 2026).

Within the confines of the reported study, the method’s significance lies in three linked claims: first, natural images exhibit power-law spectra; second, a diffusion model should respect that structure in the forward process rather than overwrite it with isotropic white noise; third, doing so produces a reverse trajectory that is closer to linear and therefore easier to sample with fewer steps. The CIFAR-10 results provide the concrete instantiation of that argument, with $P(k)\propto k^{-\alpha}$8 emerging both from spectral analysis and from empirical ablation, and with FID $P(k)\propto k^{-\alpha}$9 at $\epsilon$0 exceeding the white-noise baseline at $\epsilon$1.

In that sense, Flicker-DDPM can be understood as a principled reparameterization of diffusion noise around the spectral statistics of the target data distribution.