Step-Aware Residual-Guided Diffusion for EEG Spatial Super-Resolution
Hongjun Liu, Leyu Zhou, Zijianghao Yang, Chao Yao
TL;DR
This paper introduces SRGDiff, a novel step-aware residual-guided diffusion model that significantly improves EEG spatial super-resolution, enhancing high-density EEG reconstruction from sparse measurements for better analysis and visualization.
Contribution
The paper proposes a dynamic residual condition within a diffusion model to improve EEG super-resolution, addressing distribution shift and signal distortion issues.
Findings
Achieves up to 40% performance gain over baselines.
Effectively mitigates spatial-spectral shift in EEG signals.
Demonstrates superior results across multiple datasets and metrics.
Abstract
For real-world BCI applications, lightweight Electroencephalography (EEG) systems offer the best cost-deployment balance. However, such spatial sparsity of EEG limits spatial fidelity, hurting learning and introducing bias. EEG spatial super-resolution methods aim to recover high-density EEG signals from sparse measurements, yet is often hindered by distribution shift and signal distortion and thus reducing fidelity and usability for EEG analysis and visualization. To overcome these challenges, we introduce SRGDiff, a step-aware residual-guided diffusion model that formulates EEG spatial super-resolution as dynamic conditional generation. Our key idea is to learn a dynamic residual condition from the low-density input that predicts the step-wise temporal and spatial details to add and uses the evolving cue to steer the denoising process toward high density reconstructions. At each…
Peer Reviews
Decision·ICLR 2026 Poster
Originality. The idea of dynamically guiding each denoising step with a residual derived from the LD forward process is novel and conceptually distinct from prior approaches that rely on static conditioning. Quality. The method is comprehensively evaluated across three public datasets and multiple SR scales, covering signal-, feature-, and downstream-level tasks. Ablations demonstrate the importance of both RDM and SMM. Significance. SRGDiff consistently achieves superior performance measured
- The authors claim that SRGDiff is transferable to general SR settings, while all experiments are restricted to EEG. This should at least be discussed in more detail but ideally be experimentally verified to support the claim. - The impact of $\lambda_{res}$ and $\lambda_{SMM}$ is not studied. Adding an experiment to the ablation section where $\lambda_{res}$ and $\lambda_{SMM}$ are varied would provide insight into the sensitivity and impact of hyperparameters. - In the abstract, the claim of
- The work shows thorough experimentation by testing their approach across multiple datasets - The paper is well-writen and cleanly organized - The quantitative improvements over previous EEG SR methods are consistent and strong
Besides their meaningful application to an important domain, the proposed method largely combines existing diffusion-conditioning tricks (residual conditioning, affine modulation, step embeddings). There is little theoretical or conceptual innovation beyond standard conditional diffusion formulations. Overall, the method resembles ControlNet/T2I-Adapter-style modulation, rebranded for EEG. For the ICLR main track, this feels closer to an applied engineering paper than a new machine learning cont
The paper introduces EEG spatial super-resolution to the broader ML community, addressing a domain problem that remains underexplored. The use of deep generative modeling (transformer/diffusion) provides a fresh analytical perspective and helps re-interpret a traditionally biomedical task in a form understandable to the ICLR community. The overall framework and evaluation pipeline are consistent with prior EEG SR baselines (e.g., ESTformer, STAD), allowing fair comparison and reproducibility.
1. The authors continue using NMSE, PCC, and SNR as the main evaluation metrics. These are inherited from previous EEG works but lack a theoretical justification for super-resolution tasks in the ML setting. Given the ICLR context, this presents an opportunity to re-examine whether these metrics are conceptually valid critically. For instance, does high-density (HD) EEG necessarily yield higher SNR values than low-density EEG? If not, why do the reported gains occur? 2. EEG signals vary signifi
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsEEG and Brain-Computer Interfaces · Functional Brain Connectivity Studies · Neural dynamics and brain function