Skip to content

Step-Aware Residual-Guided Diffusion for EEG Spatial Super-Resolution

Conference: ICLR 2026 arXiv: 2510.19166 Code: GitHub Area: Diffusion Models / EEG Signals / Super-Resolution Keywords: EEG super-resolution, residual-guided diffusion, step-aware modulation, brain-computer interface, conditional generation

TL;DR

This paper proposes SRGDiff, a step-aware residual-guided diffusion model that reformulates EEG spatial super-resolution as a dynamic conditional generation task, achieving high-fidelity reconstruction via per-step residual direction correction and timestep-dependent affine modulation.

Background & Motivation

EEG (electroencephalography) is a non-invasive brain activity monitoring technique widely used in brain-computer interfaces, epilepsy diagnosis, and emotion recognition. However:

Limited spatial resolution: High-density (HD) systems are costly and inconvenient to wear; low-density (LD) systems (8–16 electrodes) are practical but suffer from severe spatial aliasing.

Limitations of existing super-resolution methods: - Direct feature mapping methods (CNN/Transformer) oversimplify nonlinear dependencies, producing overly smooth results - GAN-based methods require large amounts of data and computation - Static conditioning strategies in diffusion models lead to a trade-off between distributional shift and distortion

Core challenge: The tension between fidelity (generating HD-like content) and consistency (agreement with LD observations).

Method

Problem Formulation

  • Low-density EEG: \(X^L \in \mathbb{R}^{C_L \times Length}\)
  • High-density EEG: \(X^H \in \mathbb{R}^{C_H \times Length}\), \(C_H > C_L\)
  • Objective: Recover \(X^H\) from \(X^L\)

1. Latent Diffusion Backbone

  • A pre-trained VAE encoder-decoder is trained on HD EEG data
  • The loss comprises a reconstruction term, an STFT spectral fidelity term, and a KL regularization term
  • VAE parameters are frozen after convergence

2. Residual Direction Module (RDM)

Core Idea: Learn to predict residual directions from LD inputs along the forward noising trajectory, serving as a per-step correction signal.

  • Residual target: \(\delta z_t = z_0 - z_t\) (difference between HD latent and noised latent)
  • Lightweight convolutional predictor \(R_\phi\) predicts the residual: \(Res_t = R_\phi(\tau(t), c)\)
  • Residual loss: \(\mathcal{L}_{res} = \sum_t \|Res_t - \delta z_t\|_2^2\)
  • Additive fusion: \(\hat{z}_t^{RDM} = \text{LayerNorm}(\hat{z}_t) + Res_t\)

3. Step-Aware Modulation Module (SMM)

Controls the influence of residual conditioning on denoising:

  • Fuses LD features and timestep embeddings: \(\widetilde{h}_t = \sigma_t h_t + (1-\sigma_t) e_t\)
  • Predicts channel-wise scale and bias: \(\hat{z}_t^{SMM} = \gamma_t \odot \hat{z}_t^{RDM} + \beta_t^c\)
  • Weight \(\sigma_t\) decays linearly with the timestep

4. Two-Stage Training

Stage 1: VAE pre-training (HD data only) Stage 2: Residual-guided latent diffusion

\[\mathcal{L}_{\text{Stage 2}} = \mathbb{E}[\|\epsilon - \epsilon_\theta(z_t, t, c)\|_2^2] + \lambda_{res}\sum_t\|R_\varphi(c,t) - (z_0 - z_t)\|_2^2 + \lambda_{SMM}(\|\gamma_t - 1\|_2^2 + \|\beta_t\|_2^2)\]

Key Experimental Results

Datasets

  • SEED: 62 channels, 1000 Hz, emotion recognition (positive/neutral/negative)
  • SEED-IV: 62 channels, 4 emotion classes
  • Localize-MI: 256 channels, 8000 Hz, epileptic stimulation

Main Results (Localize-MI)

Method 2× SNR 4× SNR 8× SNR 16× SNR
SaSDim 5.74 4.38 3.55 2.77
SADI 5.75 4.37 3.55 2.89
RDPI 5.73 — — —
ESTformer baseline baseline baseline baseline
STAD baseline+ baseline+ baseline+ baseline+
SRGDiff best best best best

Key Findings

  • Relative SNR improvement of approximately 75% on the most challenging 8× setting
  • Significant improvements in both topographic map visualization and EEG-FID metrics
  • Effectively mitigates spatial-spectral shift between low-density and high-density recordings

Three-Level Evaluation Protocol

  1. Signal level: SNR, NMSE, PCC (temporal consistency, spectral fidelity, spatial topology)
  2. Feature level: EEG-FID (representation quality)
  3. Downstream level: Classification accuracy

Ablation Study

Component SNR Change
w/o RDM Significant drop
w/o SMM Moderate drop
Static conditioning (concatenation / cross-attention) Below dynamic conditioning
Full SRGDiff Best

Highlights & Insights

  1. Dynamic conditional generation paradigm: Couples the LD forward noising trajectory with the HD reverse denoising trajectory
  2. Residual guidance direction: Unlike static conditioning, provides directional correction at each step
  3. Comprehensive three-level evaluation: Goes beyond pointwise error to cover signal, feature, and downstream task dimensions
  4. Robustness across datasets and upscaling factors

Limitations & Future Work

  1. Requires VAE pre-training and two-stage training, resulting in a relatively complex pipeline
  2. Relies on spatial correspondence between LD and HD channels
  3. Diffusion model inference speed limits real-time BCI applications
  4. Accuracy at extreme super-resolution factors (e.g., 16×) still has room for improvement
  • EEG super-resolution: EEGSR-GAN, ESTformer, STAD, DDPM-EEG
  • Time-series diffusion: Diffusion-TS, SaSDim, SADI
  • Residual diffusion: PET-MRI residual synthesis, event-driven video residual reconstruction

Rating

  • Novelty: ⭐⭐⭐⭐ — Residual guidance combined with step-aware modulation is novel in the EEG domain
  • Value: ⭐⭐⭐⭐ — Significant practical value for low-cost BCI devices
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ — Three-level evaluation protocol is comprehensively designed
  • Writing Quality: ⭐⭐⭐⭐ — Method description is clear with sufficient ablation analysis