Residual-Guided DDPM: Correction-Driven Diffusion

• Residual-guided DDPMs are diffusion models that incorporate a structured residual term to explicitly model corrections between degraded and target images.
• The approach modifies both the forward diffusion and reverse denoising steps, enabling accelerated sampling and improved restoration performance.
• Applied to MRI motion correction, the method achieves state-of-the-art speed and fidelity, outperforming traditional diffusion and GAN-based techniques.

Residual-guided denoising diffusion probabilistic models are a family of diffusion formulations in which the degradation residual, an initial restoration, or a prior residual term is incorporated into the diffusion pipeline rather than treating restoration as unconditional denoising from pure Gaussian noise alone. In this paradigm, the reverse process is guided by a structured discrepancy between a target and a degraded observation, or by the residual of a coarse predictor. Within MRI motion correction, Res-MoCoDiff is an efficient denoising diffusion probabilistic model tailored for MRI motion artifact correction; it introduces a residual error shifting mechanism in the forward diffusion process, aligns the noise distribution with motion-corrupted data, and enables an efficient four-step reverse diffusion (Safari et al., 6 May 2025). Related residual-guided formulations appear in image restoration through prior residual injection (Shi et al., 2023), in conditional restoration through residual learning around an initial-guidance U-Net (Zhang et al., 2023), in text-to-speech through residual spectrogram refinement (Chen et al., 2022), and in ensemble learning through symmetric residual cancellation between diffusion and end-to-end predictors (Zhenning et al., 2023).

1. Residual guidance as a modeling principle

The central object in residual-guided DDPMs is an explicitly defined discrepancy term. In Res-MoCoDiff, for a motion-free image $x \in \mathbb{R}^n$ and its motion-corrupted counterpart $y$, the residual map is

$r = y - x.$

The forward chain then shifts samples in the direction of $r$ while injecting Gaussian noise, so that the terminal distribution is centered near the motion-corrupted image rather than near zero (Safari et al., 6 May 2025). In Resfusion, the degraded image $\hat x_0$ and ground truth $x_0$ define a prior residual

$R = \hat x_0 - x_0,$

and the forward process fuses $R$ into the chain so that the reverse process starts directly from noisy degraded images (Shi et al., 2023).

A closely related formulation appears in conditional image restoration, where the degraded input $x$ is first mapped to a coarse “initial restoration” $G_0 = u_\phi(x)$, and the diffusion model is trained on the residual

$y$0

with $y$1 the clean target (Zhang et al., 2023). In text-to-speech, ResGrad defines the residual between the mel-spectrogram predicted by a base non-iterative TTS model and the ground-truth speech, again shifting the generation target from the full signal to a correction term (Chen et al., 2022). In segmentation-oriented ensemble learning, ResEnsemble-DDPM defines the residual error of a pretrained end-to-end model and trains the diffusion branch to recover a symmetric, negated residual target before averaging outputs (Zhenning et al., 2023).

This suggests that “residual-guided DDPM” is not a single algorithm but a design pattern: the diffusion process is specialized to model a structured correction term that already encodes task-specific low-frequency information, deterministic structure, or model bias.

2. Forward diffusion with residual injection

In Res-MoCoDiff, the forward diffusion is defined by a monotonically increasing sequence $y$2, with $y$3. Each step perturbs the image by shifting in the direction of the residual and adding Gaussian noise:

$y$4

Marginalizing over the chain yields

$y$5

so that at $y$6 one arrives at

$y$7

To control $y$8 non-uniformly, Res-MoCoDiff adopts the geometric scheduler of Yue et al.:

$y$9

with $r = y - x.$0 governing the rate at which residual error is accumulated, and with $r = y - x.$1, $r = y - x.$2, and $r = y - x.$3 so that early steps remain close to $r = y - x.$4 (Safari et al., 6 May 2025).

Resfusion adopts an analogous but not identical strategy. Its forward kernel is

$r = y - x.$5

which unrolls to

$r = y - x.$6

The paper then defines a weighted residual-noise target, “resnoise,” that mixes the original noise and a scaled version of $r = y - x.$7 (Shi et al., 2023).

A common misconception is that residual guidance is equivalent to ordinary conditional diffusion with a degraded image appended at the input. The image-restoration literature represented here explicitly distinguishes the two cases: traditional diffusion-based image restoration methods utilize degraded images as conditional input without modifying the original denoising diffusion process, whereas residual-guided variants alter the forward process or the denoising target itself (Shi et al., 2023).

3. Reverse denoising and accelerated sampling

Res-MoCoDiff is defined around a four-step reverse diffusion rather than the hundreds of steps in conventional DDPMs. The reverse chain is

$r = y - x.$8

with Gaussian transitions

$r = y - x.$9

where the true posterior variance is fixed and input-independent:

$r$0

The mean is parameterized by the network prediction $r$1 of the clean image:

$r$2

Starting from

$r$3

each reverse step uses

$r$4

with $r$5 and $r$6 if $r$7 (Safari et al., 6 May 2025).

In Resfusion, acceleration is justified by a smooth equivalence transformation. Rewriting the forward sample gives

$r$8

The coefficient of $r$9 passes through zero when $\hat x_0$0, leading to the “optimal acceleration step”

$\hat x_0$1

after which both training and inference are truncated to $\hat x_0$2 (Shi et al., 2023).

Taken together, these formulations indicate that acceleration in residual-guided diffusion does not merely come from generic sampler reduction. A plausible implication is that the structured residual term changes the initialization and target geometry enough that a short reverse chain becomes viable without starting from pure Gaussian white noise.

4. Architectural realizations

Res-MoCoDiff implements the denoiser $\hat x_0$3 as a U-Net in which the usual self-attention modules are replaced by hierarchical Swin-Transformer blocks. In each down- and up-sampling stage, local windows of size $\hat x_0$4 shift between layers, enabling multi-scale feature aggregation via the U-Net’s encoder-decoder and skip links, local and cross-window self-attention for long-range context, and parameter-efficient modeling of fine details at multiple resolutions. Time-step embeddings and the motion-corrupted image $\hat x_0$5 are concatenated with $\hat x_0$6 as additional channels in the input (Safari et al., 6 May 2025).

The conditional restoration framework of (Zhang et al., 2023) uses a different residual-guided architecture. It combines a lightweight initial-guidance U-Net with a continuous-noise DDPM-style U-Net whose residual blocks are replaced by a Basic Module and a Conditional Integration Module. The Conditional Integration Module upsamples $\hat x_0$7, embeds scalar conditions such as $\hat x_0$8 and degradation type, and uses an Adaptive Kernel Guidance Module to construct per-location fused kernels:

$\hat x_0$9

Dynamic kernels make the conditioning spatially adaptive, so that different spatial locations can be denoised differently under the guidance $x_0$0 (Zhang et al., 2023).

Other residual-guided variants use lighter backbones. Resfusion uses a single standard U-Net, exactly as in RDDM, with input channels $x_0$1 and output equal to the resnoise predictor (Shi et al., 2023). ResGrad uses a lightweight U-Net similar to Grad-TTS with approximately $x_0$2 M parameters and conditions each block on the base mel estimate $x_0$3 (Chen et al., 2022).

5. Training objectives and prediction targets

Res-MoCoDiff minimizes a summed reconstruction error over random $x_0$4:

$x_0$5

Both terms are weighted equally, where $x_0$6 promotes global fidelity and $x_0$7 sharpens edges and reduces pixel-level bias (Safari et al., 6 May 2025).

By contrast, Resfusion keeps an $x_0$8 prediction loss but changes the target from pure Gaussian noise to resnoise:

$x_0$9

The goal is for a single predictor to learn both to remove Gaussian noise and to subtract the prior residual in the correct proportions (Shi et al., 2023).

The unified conditional restoration framework jointly trains a noise predictor and a residual correction branch under a modified noise-prediction objective that incorporates residual modeling. Its direct form includes a second term weighted by $R = \hat x_0 - x_0,$0 for residual fidelity, while also noting that the second term can be absorbed into the first by predicting $R = \hat x_0 - x_0,$1 correctly (Zhang et al., 2023). ResGrad reverts to the simplified $R = \hat x_0 - x_0,$2-prediction loss over the residual spectrogram:

$R = \hat x_0 - x_0,$3

with conditioning on the base TTS estimate (Chen et al., 2022).

These losses show two distinct residual-guided strategies. One predicts the clean target directly from a residual-shifted state, as in Res-MoCoDiff. The other predicts a noise-like quantity defined in the residual domain, as in Resfusion and ResGrad.

6. MRI motion correction via Res-MoCoDiff

In MRI motion correction, Res-MoCoDiff is evaluated on an in-silico dataset and an in-vivo dataset. The in-silico data comprise 580 T1-w IXI brain scans, split into 480 train and 100 test, with motion simulated by perturbing k-space lines. Minor motion uses 7 lines with rotation $R = \hat x_0 - x_0,$4 and shift $R = \hat x_0 - x_0,$5 mm; moderate motion uses 10 lines; heavy motion uses 15 lines. The in-vivo data are from the MR-ART dataset with 148 subjects and include ground truth, level-1, and level-2 motion images, rigidly registered. Metrics are PSNR, SSIM, and NMSE computed via the PIQ library. Baselines are Pix2pix, CycleGAN, and MT-DDPM, described as a vision-transformer DDPM (Safari et al., 6 May 2025).

On the in-silico test set, Res-MoCoDiff reports the following values.

Motion severity	PSNR	SSIM / NMSE
Minor distortion	41.91 ± 2.94 dB	0.99 ± 0.00 / 0.10 ± 0.09%
Moderate	37.97 ± 2.39 dB	0.98 ± 0.01 / 0.24 ± 0.16%
Heavy	34.15 ± 2.42 dB	0.96 ± 0.01 / 0.58 ± 0.40%

These results consistently outperform Pix2pix, CycleGAN, and MT-DDPM in SSIM and NMSE, while achieving competitive PSNR. On in-vivo MR-ART, the reported values are 30.40 ± 2.90 dB, 0.92 ± 0.05, and 1.71 ± 1.49% at level 1, and 29.63 ± 2.97 dB, 0.91 ± 0.05, and 2.07 ± 1.79% at level 2, with reported relative changes of $R = \hat x_0 - x_0,$6, $R = \hat x_0 - x_0,$7, and $R = \hat x_0 - x_0,$8 for level 1, and $R = \hat x_0 - x_0,$9, $R$0, and $R$1 for level 2 (Safari et al., 6 May 2025).

Inference time is 0.37 s per two-slice batch, versus approximately 101.74 s for MT-DDPM. The paper also characterizes this as a four-step sampling time of approximately 0.37 s for two slices and a $R$2 speed-up over standard DDPM inference at approximately 101.7 s. Its summary describes the resulting performance as state-of-the-art motion artifact removal with clinical-scale speed, high structural fidelity, low NMSE, and PSNR up to 41.9 dB on simulated data (Safari et al., 6 May 2025).

7. Relation to adjacent residual-guided formulations, limitations, and interpretive issues

Residual guidance appears in several adjacent but non-identical forms. In ResGrad, the diffusion model is employed in the inference process of the existing TTS model in a plug-and-play way, without re-training this model, and the paper attributes the speed-quality trade-off to changing the generation target from ground-truth mel-spectrogram to the residual (Chen et al., 2022). In ResEnsemble-DDPM, the diffusion branch and a frozen pretrained end-to-end model are combined through a symmetric construction: the end-to-end model predicts $R$3, the diffusion model is trained toward $R$4, and the final output is the average

$R$5

The paper states that averaging exactly cancels the error $R$6 in the idealized symmetric case, while also noting that full quantitative results and ablations are not yet published and that the theoretical analysis is informal (Zhenning et al., 2023).

A second interpretive issue concerns whether residual guidance primarily modifies the forward process, the reverse conditioning, or the prediction target. The literature in this set contains all three possibilities. Res-MoCoDiff uses explicit residual injection into the forward process and predicts the clean image under conditioning on the corrupted observation (Safari et al., 6 May 2025). Resfusion injects a prior residual into the forward chain and predicts resnoise (Shi et al., 2023). The unified conditional framework learns the residual of an initial guidance and adds a residual correction term during denoising (Zhang et al., 2023). This suggests that “residual-guided DDPM” should be understood as a structural family organized around residualized state variables or targets, rather than a single canonical parameterization.

A third issue is whether residual guidance necessarily requires an auxiliary coarse predictor. The surveyed formulations do not support a single answer. ResGrad depends on an existing TTS model, and ResEnsemble-DDPM depends on a frozen pretrained end-to-end model (Chen et al., 2022, Zhenning et al., 2023). By contrast, Res-MoCoDiff is formulated directly from paired motion-free and motion-corrupted MRI data through $R$7, and Resfusion starts from the degraded input and its residual to the target without requiring a separate ensemble learner (Safari et al., 6 May 2025, Shi et al., 2023).

Across these variants, the consistent theme is that residual guidance narrows the denoising objective to a structured correction problem. The exact benefit depends on the task-specific way that residuals are defined, injected, and predicted.