Perturbed Consistency Diffusion Model

• The paper introduces a system where adaptive illumination correction is integrated as an intermediate diffusion state to guide the image restoration process.
• The approach enforces consistency by anchoring a perturbed state to the reverse diffusion trajectory, combining denoising and trajectory regularization.
• This method demonstrates that coupling conditional reverse diffusion with perturbation yields robust restoration without needing paired ground truths.

Searching arXiv for papers on perturbed consistency diffusion models and related consistency-model literature. Perturbed Consistency Diffusion Model (PCDM) most concretely denotes, in recent image-restoration literature, a diffusion-based reconstruction module that treats a pre-corrected observation as an intermediate noisy state and then enforces consistency between that perturbed state and the model’s reconstructed trajectory. In "ZeroIDIR: Zero-Reference Illumination Degradation Image Restoration with Perturbed Consistency Diffusion Models" (Jiang et al., 12 May 2026), this idea is instantiated for zero-reference illumination degradation image restoration by coupling adaptive illumination correction with conditional reverse diffusion. Closely related work uses "perturbed consistency" in a more theoretical sense: a learned consistency model is followed by an additional Gaussian perturbation step to obtain stronger total-variation guarantees (Chen et al., 6 May 2025). Across these usages, the common motif is trajectory anchoring under perturbation rather than unrestricted denoising alone.

1. Terminological scope and defining idea

The phrase "perturbed consistency diffusion model" is not used in a single uniform way across the literature represented here. In ZeroIDIR, it refers to a conditional diffusion model trained on degraded images only, where the adaptive gamma-corrected image is reinterpreted as a diffusion state $x_{t^*}$ and the reverse model is constrained so that its reconstructed sample, when pushed forward again, remains consistent with that intermediate input (Jiang et al., 12 May 2026). In the consistency-model theory of Chen et al., the "perturbed" component is a final Gaussian smoothing step appended to one-step or multistep sampling, introduced to obtain total-variation convergence under log-density smoothness assumptions (Chen et al., 6 May 2025). In ConsistencyDet, consistency is instead the self-consistency of a detector that maps perturbed boxes from adjacent times back to the same clean boxes in a few denoising steps (Jiang et al., 2024).

Context	Perturbed object	Consistency mechanism
ZeroIDIR	illumination-corrected image treated as $x_{t^*}$	feature-space anchoring of re-diffused prediction
Consistency-model theory	output law convolved with $N(0,\sigma_\eps^2 I)$	approximate self-consistency along ODE trajectories
ConsistencyDet	bounding boxes with Gaussian noise	adjacent-time denoising to the same GT box

A plausible implication is that PCDM should be understood as a family resemblance rather than a single canonical architecture. The defining property is not merely the presence of diffusion and a consistency loss, but the imposition of compatibility between a perturbed intermediate representation and a reconstruction path derived from it.

2. ZeroIDIR formulation: intermediate-state conditioning for zero-reference restoration

ZeroIDIR proposes a zero-reference diffusion-based framework for illumination degradation image restoration that decouples the restoration process into adaptive illumination correction and diffusion-based reconstruction while being trained solely on low-quality degraded images (Jiang et al., 12 May 2026). Its core intuition is to treat the output of a lightweight adaptive gamma correction stage as if it were already partway along a standard forward-diffusion trajectory, then continue the forward diffusion by a random number of steps, and finally learn a conditional reverse-diffusion that not only denoises but also "re-anchors" its final forward-diffusion path to the intermediate input.

The diffusion notation follows the standard DDPM-style formulation. With noise levels $\{\beta_1,\dots,\beta_T\}$ and $\bar\alpha_t \triangleq \prod_{s=1}^t (1-\beta_s)$, the forward process is

$$q(x_t\mid x_{t-1})=\mathcal N\!\Bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1},\beta_t I\Bigr)$$

and therefore

$$q(x_t\mid x_0)=\mathcal N\!\Bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\Bigr).$$

In ZeroIDIR, the true well-illuminated target $x_0$ is never observed. Instead, the adaptive gamma-correction module produces an image $I_d'$, and the model reinterprets this image as the diffusion state at an unknown intermediate time $t^*$, written as $x_{t^*}$0. Training samples for the diffusion model are then generated by continuing the forward process from that state for an extra $x_{t^*}$1 steps, with $x_{t^*}$2 chosen at random.

The reverse process is parameterized in the usual Gaussian form with a noise network $x_{t^*}$3, conditioned on $x_{t^*}$4. In the reported implementation, $x_{t^*}$5 is a U-Net that takes as input both the noisy image $x_{t^*}$6 and the spatially adaptive illumination-corrected map $x_{t^*}$7, typically by concatenation in the first layer or via cross-attention. Because $x_{t^*}$8 already has most low-frequency illumination information corrected, the U-Net can focus its capacity on recovering fine textures and suppressing noise. This suggests that ZeroIDIR uses perturbation not as a generic augmentation device, but as an explicit latent-state reinterpretation that turns illumination correction into a conditioning signal for detail reconstruction.

3. Two-stage architecture: AGCM followed by PCDM

ZeroIDIR is organized as a two-stage system. The first stage is the adaptive gamma correction module (AGCM). Given a degraded image $x_{t^*}$9, the method performs Retinex decomposition $N(0,\sigma_\eps^2 I)$0, predicts spatial gamma maps $N(0,\sigma_\eps^2 I)$1 and weights $N(0,\sigma_\eps^2 I)$2, forms

$N(0,\sigma_\eps^2 I)$3

and then reconstructs the illumination-corrected image as

$N(0,\sigma_\eps^2 I)$4

The purpose of this stage is to generate illumination-corrected only representations that mitigate exposure bias and serve as reliable inputs for subsequent diffusion processes. A histogram-guided illumination correction loss is introduced to regularize the corrected illumination distribution toward that of natural scenes, and the AGCM is trained by minimizing $N(0,\sigma_\eps^2 I)$5, described as histogram-guided KL loss, exposure control, and edge-aware TV (Jiang et al., 12 May 2026).

The second stage is the PCDM. For each batch, the model samples a degraded image $N(0,\sigma_\eps^2 I)$6, computes $N(0,\sigma_\eps^2 I)$7 via the fixed AGCM, samples $N(0,\sigma_\eps^2 I)$8 and $N(0,\sigma_\eps^2 I)$9, sets $\{\beta_1,\dots,\beta_T\}$0, draws $\{\beta_1,\dots,\beta_T\}$1, forms the noisy state $\{\beta_1,\dots,\beta_T\}$2 from $\{\beta_1,\dots,\beta_T\}$3, computes the denoising and perturbed-consistency losses, and updates $\{\beta_1,\dots,\beta_T\}$4 by descending $\{\beta_1,\dots,\beta_T\}$5. The implementation highlights additionally specify $\{\beta_1,\dots,\beta_T\}$6 and $\{\beta_1,\dots,\beta_T\}$7, with $\{\beta_1,\dots,\beta_T\}$8 diffusion steps and a linear $\{\beta_1,\dots,\beta_T\}$9-schedule.

At inference, a single degraded image $\bar\alpha_t \triangleq \prod_{s=1}^t (1-\beta_s)$0 is first converted into $\bar\alpha_t \triangleq \prod_{s=1}^t (1-\beta_s)$1 with the trained AGCM, and $\bar\alpha_t \triangleq \prod_{s=1}^t (1-\beta_s)$2 is set. Reverse diffusion then proceeds from $\bar\alpha_t \triangleq \prod_{s=1}^t (1-\beta_s)$3, predicting $\bar\alpha_t \triangleq \prod_{s=1}^t (1-\beta_s)$4 and sampling $\bar\alpha_t \triangleq \prod_{s=1}^t (1-\beta_s)$5 from the corresponding Gaussian transition. The final denoised image $\bar\alpha_t \triangleq \prod_{s=1}^t (1-\beta_s)$6 is obtained by one more posterior step from $\bar\alpha_t \triangleq \prod_{s=1}^t (1-\beta_s)$7. In the implementation summary, ancestral sampling is performed in 20 steps at test time (Jiang et al., 12 May 2026).

4. Loss design: denoising and perturbed diffusion consistency

The PCDM training objective combines a standard denoising term with a trajectory-anchoring consistency term. The denoising loss is

$\bar\alpha_t \triangleq \prod_{s=1}^t (1-\beta_s)$8

Here $\bar\alpha_t \triangleq \prod_{s=1}^t (1-\beta_s)$9 is formed from $$q(x_t\mid x_{t-1})=\mathcal N\!\Bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1},\beta_t I\Bigr)$$0. Minimizing $$q(x_t\mid x_{t-1})=\mathcal N\!\Bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1},\beta_t I\Bigr)$$1 makes $$q(x_t\mid x_{t-1})=\mathcal N\!\Bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1},\beta_t I\Bigr)$$2 predict the true Gaussian noise used in the forward process.

The second term is the perturbed diffusion consistency loss,

$$q(x_t\mid x_{t-1})=\mathcal N\!\Bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1},\beta_t I\Bigr)$$3

where $$q(x_t\mid x_{t-1})=\mathcal N\!\Bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1},\beta_t I\Bigr)$$4 is a fixed perceptual feature extractor, exemplified by VGG-16 conv-features, $$q(x_t\mid x_{t-1})=\mathcal N\!\Bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1},\beta_t I\Bigr)$$5 is the cleaned output inferred from $$q(x_t\mid x_{t-1})=\mathcal N\!\Bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1},\beta_t I\Bigr)$$6, and $$q(x_t\mid x_{t-1})=\mathcal N\!\Bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1},\beta_t I\Bigr)$$7 is obtained by pushing $$q(x_t\mid x_{t-1})=\mathcal N\!\Bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1},\beta_t I\Bigr)$$8 forward again for $$q(x_t\mid x_{t-1})=\mathcal N\!\Bigl(x_t;\sqrt{1-\beta_t}\,x_{t-1},\beta_t I\Bigr)$$9 steps. Intuitively, $$q(x_t\mid x_0)=\mathcal N\!\Bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\Bigr).$$0 is what one would obtain by taking the model’s cleaned $$q(x_t\mid x_0)=\mathcal N\!\Bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\Bigr).$$1 and re-diffusing it back to the intermediate time; the loss forces that re-diffused image to match the original $$q(x_t\mid x_0)=\mathcal N\!\Bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\Bigr).$$2 in feature space, thereby anchoring the model’s generation path to the adaptive-gamma-corrected input.

The overall stage-two loss is $$q(x_t\mid x_0)=\mathcal N\!\Bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\Bigr).$$3, with $$q(x_t\mid x_0)=\mathcal N\!\Bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\Bigr).$$4 in the implementation highlights. Stage-one hyperparameters are also specified: $$q(x_t\mid x_0)=\mathcal N\!\Bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\Bigr).$$5, $$q(x_t\mid x_0)=\mathcal N\!\Bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\Bigr).$$6, exposure target $$q(x_t\mid x_0)=\mathcal N\!\Bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\Bigr).$$7, and TV weight $$q(x_t\mid x_0)=\mathcal N\!\Bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\Bigr).$$8. Histogram guidance uses $$q(x_t\mid x_0)=\mathcal N\!\Bigl(x_t;\sqrt{\bar\alpha_t}\,x_0,(1-\bar\alpha_t)I\Bigr).$$9, where the empirical histogram $x_0$0 is aggregated from $x_0$1 K well-lit images (Jiang et al., 12 May 2026).

A common misconception is to equate this loss design with the direct self-consistency objective of the consistency-model literature. ZeroIDIR still trains a U-Net noise predictor inside a DDPM-style reverse process; its "consistency" is a constraint on the compatibility between the restored sample and the perturbed intermediate state. This suggests a narrower and more restoration-specific use of the term.

5. Relation to consistency models with perturbative sampling guarantees

Chen et al. study consistency models under general data assumptions and provide a formal exposition of a perturbed consistency sampler (Chen et al., 6 May 2025). The setup begins with a forward diffusion SDE on $x_0$2,

$x_0$3

and its probability-flow ODE with the same marginals. The ground-truth consistency function is $x_0$4, and exact self-consistency means that if $x_0$5 and $x_0$6 lie on the same ODE trajectory, then $x_0$7. In practice, an approximate consistency model $x_0$8 is trained so that the training-time self-consistency loss is small on a discrete mesh.

The analysis requires only one of three conditions on the data distribution: bounded support, exponential tail decay, or log-density smoothness. Under these assumptions, the paper studies one-step sampling, multistep sampling, and a perturbed sampler in which a final Gaussian smoothing step is inserted:

$x_0$9

This perturbation is introduced because, when the target distribution is smooth, the generated samples can then be controlled in total variation distance. The paper provides a Wasserstein-$I_d'$0 bound for the $I_d'$1-step sampler under bounded support, an analogous result up to an extra $I_d'$2 term under exponential tails, and a total-variation bound for the smoothed output under log-density $I_d'$3-smoothness. Two case studies are given: the Ornstein–Uhlenbeck SDE and the Variance-Exploding SDE. For OU, a two-step schedule yields a two-step error about half the one-step error when $I_d'$4. For VE, a halving schedule yields a rate of order $I_d'$5.

This theory clarifies an important distinction. In Chen et al., "perturbed consistency" refers to smoothing the output law of a learned consistency function. In ZeroIDIR, the perturbation is an intermediate image state derived from adaptive gamma correction, and the consistency condition is imposed by matching a re-diffused reconstruction back to that state. The mechanisms are different, but both use perturbation to stabilize or certify a reconstruction path.

6. Adjacent vision uses and interpretive boundaries

ConsistencyDet is an important neighboring example because it frames object detection as a denoising process on perturbed bounding boxes and explicitly emphasizes the self-consistency feature of the consistency model (Jiang et al., 2024). Ground-truth boxes $I_d'$6 are perturbed by Gaussian noise with scale $I_d'$7, and the model $I_d'$8 is trained so that adjacent-time boxes $I_d'$9 and $t^*$0 both denoise back to the same GT box. Inference starts from random Gaussian proposals and iteratively refines them in a few denoising steps. The reported ablations show that 2–10 steps suffice, with AP saturating by 4–10 steps, and the ResNet-50/FPN configuration reaches $t^*$1 AP on COCO 2017 val in 2 steps, compared with $t^*$2 AP for DiffusionDet in 4 steps; on a single RTX3080, the 2-step model runs at $t^*$3 FPS versus $t^*$4 FPS for DiffusionDet with 20 steps.

A further boundary case appears in semi-supervised ultrasound segmentation. MGCC uses images generated by a Latent Diffusion Model as unlabeled data and imposes output consistency between a main decoder and auxiliary decoders under multi-level global context perturbations such as F-Noise, F-Drop, and Dropout (Tang et al., 2023). This is a different construction again: diffusion is used for synthetic data generation, while consistency is a decoder-level regularizer rather than a diffusion-trajectory anchor.

The literature in view therefore suggests that "perturbed consistency diffusion model" should not be used as a blanket label for all diffusion systems with perturbation and consistency terms. In the strict ZeroIDIR sense, it designates a restoration pipeline in which illumination correction is front-loaded into AGCM, the corrected image is treated as a noisy intermediate state for diffusion, and a perturbed-consistency loss anchors the reverse process to that state. ZeroIDIR reports that this design outperforms state-of-the-art unsupervised competitors and is comparable to supervised methods while being more generalizable to various scenes, and that it can produce restored images that have natural exposure, suppress noise, and recover fine details without ever seeing paired ground truths during diffusion training (Jiang et al., 12 May 2026).