Multi-Frequency Reconstruction Diffusion Model

• The MFRD model is a diffusion-driven approach that decomposes data into multiple frequency bands to enhance reconstruction fidelity.
• It employs analytical and learned frequency operators like Fourier, wavelet, and VMD to optimize score estimation across distinct subbands.
• Empirical results demonstrate improved convergence, enhanced high-frequency recovery, and reduction in reconstruction errors across applications.

The Multi-Frequency-Reconstruction-based Diffusion (MFRD) model denotes a class of diffusion-driven generative and restoration architectures that utilize explicit multi-frequency representations throughout the forward and reverse modeling pipelines. MFRD approaches aim to improve the recovery of high-frequency content and accelerate convergence in inverse problems and generative modeling—most notably, in MRI reconstruction, image super-resolution, generic image restoration, and time series forecasting. Central to these methods is the extraction or manipulation of frequency subbands, allowing separate or coordinated score models, frequency-scheduled guidance, or multiscale denoising to control the information flow and error propagation at each stage of the diffusion process.

1. Multi-Frequency Principle and Problem Setting

MFRD systematically decomposes target data into multiple frequency bands—either analytically using transforms (e.g., Fourier, wavelet packets, Variational Mode Decomposition), or by learning frequency-selective operators in k-space or image space. The rationale is rooted in two observations: first, different frequency bands carry complementary information (e.g., global structure vs. fine texture); second, forward and reverse diffusion processes can benefit from frequency-specific modeling, both in terms of fidelity and computational efficiency.

For MRI reconstruction, MFRD operates directly on k-space data, applying operators such as frequency-weighting ($W$) and frequency-domain masking ($M$) to partition and emphasize subbands (Guan et al., 2023). In time series forecasting, the original signal is decomposed into intrinsic mode functions (IMFs) via VMD, revealing hidden periodic or aperiodic patterns (Dong et al., 10 Jan 2026). Super-resolution instantiations deploy wavelet packet transforms to organize the reconstruction as a chain of partial targets with incrementally wider frequency bandwidth (Wang et al., 2024).

2. Model Architectures and Frequency Decomposition

Different MFRD variants implement the multi-frequency paradigm via distinct architectural and operational choices:

• MRI Reconstruction (CM-DM): The CM-DM (Correlated and Multi-frequency Diffusion Modeling) architecture applies two main decomposition strategies: "Weight-K-Space" (soft high-frequency emphasis via $k^w = W\odot k$) and "Mask-K-Space" (hard removal of low frequencies via $k^m = (I-M)\odot k$). Each filtered subband is passed to separate score-based generative models. Serial or parallel fusion merges their predictions in the reverse diffusion chain (Guan et al., 2023).
• Time Series Forecasting: For short-term electricity load forecasting, MFRD first applies VMD to split a signal $x(t)$ into summands $u_k(t)$ centered at different frequencies. The combined input to the denoising network is an augmented feature stack including the raw signal and all IMFs. The denoising backbone fuses a residual LSTM (for temporal modeling) and a Transformer (capturing long-range dependencies and contextual integration) (Dong et al., 10 Jan 2026).
• Super Resolution (FDDiff): FDDiff decomposes high-frequency content into a chain of intermediate targets using wavelet packet decomposition. The reverse diffusion process fills in missing high-frequency detail progressively, each inference sub-task operating at a controlled spatial and spectral scale. The denoising network is a multi-state U-Net backbone, with scale-specific heads and soft parameter-sharing to accommodate transitions across spectral bands (Wang et al., 2024).
• Diffusion Guidance (FGPS): Frequency-Guided Posterior Sampling (FGPS) introduces a frequency curriculum, dynamically ramping up the frequency content incorporated from the measurements over the sampling trajectory. At early diffusion steps, only low-frequency information guides the estimation, with higher frequencies progressively introduced as the reverse process proceeds (Thaker et al., 2024).

3. Diffusion Training and Frequency-Selective Losses

MFRD leverages the standard Denoising Diffusion Probabilistic Model (DDPM) framework for forward and reverse processes, extended to handle frequency-partitioned data.

• Forward Process: At each step, noise is injected according to a variance schedule ($\beta_t$). In MFRD, this is applied not only to the raw data but also, independently, to the various subbands or decomposed features.
• Reverse Process: Frequency-aware denoising networks ($s_\theta^w$, $s_\theta^m$, or generic $\epsilon_\theta$) are trained to invert the forward process, often using a score-matching loss on frequency-filtered representations. Some variants incorporate explicit correlation regularization to encourage agreement between multiple high-frequency score estimators (Guan et al., 2023).
• Loss Functions: Examples include standard DSM losses on subbands, joint frequency/time-domain MSEs (as in time series forecasting (Dong et al., 10 Jan 2026)), or scale- and SNR-weighted $L_1$ reconstruction errors as used in multiscale super-resolution (Wang et al., 2024).

4. Sampling, Inference Procedures, and Reconstruction

Sampling and inference procedures in MFRD are tailored to integrate information from multiple frequency branches or scales at every denoising iteration.

• In MRI, the sampling step involves running reverse updates on both $k^w$ and $k^m$, merging the results (via averaging or chaining), enforcing data consistency in acquired frequency locations, and projecting intermediate k-space estimates onto low-rank subspaces via truncated SVD on the Hankel matrix (Guan et al., 2023).
• For time series, at each reverse step the model replaces the look-back portion of each denoised sample with the observed history, ensuring deterministic handling of observational context during prediction (Dong et al., 10 Jan 2026).
• In FDDiff, the U-Net backbone's scale-specific heads manage the injection of high-frequency slices, guided by intermediate frequency targets reconstructed from partial wavelet coefficients (Wang et al., 2024).
• Frequency scheduling in FGPS adjusts the measurement's frequency mask at each sampling step, manipulating the conditional score according to an adaptive or pre-defined curriculum (Thaker et al., 2024).

5. Empirical Evaluation and Performance Gains

MFRD models consistently deliver improvements in reconstruction fidelity, convergence speed, and robustness to undersampling or noise across modalities:

Domain	Metric(s)	MFRD Result	Best Baseline	Reference
MRI (Brain)	PSNR, SSIM @ 10× Poisson	40.58 dB, 0.9387	HGGDP: 32.80 dB, 0.8986	(Guan et al., 2023)
Time Series	MAE, RMSE, MAPE, R² (AEMO-NSW)	93.38, 123.41, 1.28%, 0.992	Transformer: 109.24 MAE	(Dong et al., 10 Jan 2026)
Super-res Face	PSNR, SSIM (8×, CelebA-HQ)	24.52 dB, 0.71	IDM: 24.01 dB, 0.71	(Wang et al., 2024)
Restoration	PSNR, SSIM (high-pass FFHQ)	18.8 dB, 0.739	DPS: 6.2 dB, 0.695	(Thaker et al., 2024)

Empirical ablation studies indicate marked performance degradation if multi-frequency decomposition, specialized denoising architectures, or frequency-domain losses are omitted. In MRI, MFRD enables accurate maintenance of critical anatomical textures with markedly fewer sampling steps. In forecasting, both pointwise and spectral errors are reduced relative to Transformer, LSTM, and other neural-net baselines.

6. Theoretical Insights and Convergence Properties

The underlying justification for MFRD efficiency derives from the statistical structure of data and the mechanics of the diffusion process. Constraining learning and sampling within high-frequency subspaces improves alignment between target score distributions and added noise, decreasing the variance of reverse steps and accelerating convergence per iteration (Guan et al., 2023). In image restoration, frequency-guided scheduling limits error propagation caused by incorrect likelihood approximations at low-SNR steps (Thaker et al., 2024). Wavelet-based intermediary targets further shrink the difficulty of each diffusion subproblem by isolating a narrow frequency band, which empirically tightens generalization and reduces parameter count (Wang et al., 2024).

7. Applications and Future Directions

MFRD methods have demonstrated state-of-the-art results in:

• Highly under-sampled MRI reconstruction, yielding up to +8 dB PSNR improvement under extreme sampling rates (Guan et al., 2023).
• Short-term electricity load forecasting across real-world datasets, reducing MAE and MAPE over classical and deep learning baselines (Dong et al., 10 Jan 2026).
• Single-image super-resolution, where MFRD-based (FDDiff) models outperform GAN- and diffusion-based counterparts in both perceptual and quantitative metrics (Wang et al., 2024).
• Image restoration and deblurring, where frequency curricula prevent over-smoothing and instability during sampling (Thaker et al., 2024).

The generality of the multi-frequency diffusion principle suggests applicability to a range of inverse and generative problems in domains structured by frequency content. Future work is likely to explore adaptive or data-driven frequency band selection, tighter integration with physical measurement models, and more sophisticated coupling strategies between frequency-aware score networks.