E2I-Flow: EEG-to-Image Reconstruction via Wavelet-Temporal Cross-Attention and Conditional Flow Matching

Haowei Pan; Hwa Hui Tew; Junn Yong Loo; Fuad Noman; Hernando Ombao; Chee-Ming Ting

cs.CVNov 3, 2026

E2I-Flow: EEG-to-Image Reconstruction via Wavelet-Temporal Cross-Attention and Conditional Flow Matching

Authors: Haowei Pan, Hwa Hui Tew, Junn Yong Loo, Fuad Noman, Hernando Ombao, Chee-Ming Ting

Organizations: School of Information Technology, Monash University, Malaysia Campus · School of Electrical and Electronic Engineering, Nanyang Technological University · Statistics Program, King Abdullah University of Science and Technology

Abstract

Electroencephalography (EEG)-based visual decoding and reconstruction aims to recover visual information from non-invasive brain signals. But there are challenges due to the complex temporal-frequency characteristics of EEG signals and the difficulty of discrete denoising framework in mapping neural activity to meaningful visual representations. Here, we propose E2I-Flow, an EEG-to-image decoding and reconstruction framework that combines a novel EEG encoder for time-frequency representation learning from EEG features, and conditional flow matching for visual decoding. A stationary-wavelet frequency branch and a temporal-mixing branch are integrated through bidirectional cross-attention and adaptive fusion, with the resulting EEG representations aligned to ground-truth image embeddings using contrastive and mean-squared-error loss supervisions. The aligned EEG representation is then used as a condition for a Transformer-based flow-matching model to predict the corresponding visual embedding for image synthesis. Experimental results show strong zero-shot EEG-to-image retrieval and improved visual reconstruction performance, outperforming several state-of-the-art baselines.

Figures & tables

Figure 1: Overview of E2I-Flow. (a) EEG–image representation alignment using the time-frequency EEG encoder and a pretrained image encoder. (b) Conditional flow matching training, where the aligned EEG representation conditions the Transformer velocity model. (c) Inference through guided flow sampling followed by IP-Adapter and SDXL-Turbo for image synthesis.

Figure 2: Architecture of time-frequency EEG encoder: (a) SWT frequency branch, (b) Bidirectional cross-attention with adaptive fusion, and (c) Temporal MLP-mixing branch.

Method	Sub-01		Sub-02		Sub-03		Sub-04		Sub-05		Sub-06		Sub-07		Sub-08		Sub-09		Sub-10		Mean
Method	T1	T5	T1	T5	T1	T5	T1	T5	T1	T5	T1	T5	T1	T5	T1	T5	T1	T5	T1	T5	T1	T5
Conformer (2023) [ 20 ]	9.5	26.1	10.8	27.4	14.2	30.6	15.5	37.1	8.8	19.8	8.1	28.0	12.2	31.8	16.4	40.5	10.9	29.8	15.1	39.2	12.2	31.0
NICE (2024) [ 19 ]	20.1	47.9	19.2	36.9	21.2	53.8	22.4	50.4	14.4	31.6	26.1	56.3	26.9	55.7	38.2	67.4	23.4	53.2	29.8	58.6	24.2	51.2
ATM (2024) [ 11 ]	23.2	51.2	18.8	43.2	27.3	57.0	28.0	54.1	16.8	39.5	28.0	59.2	26.8	53.7	39.3	71.5	24.5	51.2	37.2	67.7	26.9	54.8
VDF (2025) [ 3 ]	20.4	49.0	16.5	49.0	17.5	51.0	25.5	58.0	13.0	43.0	23.0	54.5	19.5	48.0	24.0	55.0	18.0	50.0	20.5	49.5	19.8	50.7
AVDE (2026) [ 4 ]	25.0	55.2	24.1	51.0	27.5	58.6	29.8	54.7	25.4	50.3	33.5	60.3	27.4	55.2	41.7	71.3	26.1	52.1	39.5	73.0	30.0	58.2

Table 1: Within-subject 200-way zero-shot EEG-to-image retrieval performance (%) on the THINGS-EEG dataset. T1 and T5 denote Top-1 and Top-5 accuracy, respectively. Results are averaged over five random seeds.

	Low-level		High-level
Method	PixCorr $\uparrow$	SSIM $\uparrow$	Alex(2) $\uparrow$	Alex(5) $\uparrow$	Inc. $\uparrow$	CLIP $\uparrow$	SwAV $\downarrow$
ATM [ 11 ]	0.130	0.339	0.740	0.822	0.685	0.716	0.599
AVDE [ 4 ]	0.147	0.366	0.766	0.835	0.724	0.747	0.587
Ours	0.155	0.366	0.771	0.849	0.721	0.787	0.565

Table 2: Average quantitative comparison of EEG-based visual reconstruction over all 10 subjects.

Figure 3: Qualitative comparison of EEG-based image reconstruction across multiple subjects using ATM, E2I-Flow, and the corresponding ground-truth stimuli.

	Sub-01		Sub-02		Sub-09		Sub-10
Framework	T1	T5	T1	T5	T1	T5	T1	T5
SWT Frequency Only	17.0	40.0	24.5	52.5	23.5	52.5	33.5	63.0
Temporal Only	18.0	39.5	25.5	53.5	26.0	49.5	35.0	65.0
SWT + Temporal + MLP Fusion	20.0	41.5	24.5	54.5	29.0	52.5	33.5	61.5
E2I-Flow (Ours)	21.0	41.5	27.0	58.5	28.0	56.5	35.0	65.5

Table 3: E2I-Flow ablation on selected subjects. T1 and T5 denote Top-1 and Top-5 retrieval accuracy (%), respectively.

Wavelet	Level 2	Level 3	Level 4	Level 5	Level 6	Level 7
db1	48.0	43.5	47.5	43.0	44.5	44.0
db4	46.5	44.5	47.5	43.0	43.5	42.5
db8	43.0	43.5	46.5	39.5	43.0	42.0

Table 4: Hyperparameter analysis of wavelet decomposition settings. Top-1 retrieval accuracy (%) is reported for different wavelet families and decomposition levels.

EEG-to-Image Reconstruction

See all 8 papers

MindAlign: Bridging EEG, Vision, and Language for Zero-Shot Visual Decoding
May 23, 2026Zexuan Chen, Sichao Liu, Runhao Lu +4EEG-Based Visual Decoding Cross-Modal Alignment
KTH, SWeden · University of Cambridge, UK · EPFL, Switzerland +2
STAMBRIDGE: Spectral-Temporal Amplitude-aware Mid-Feature Bridge for EEG Visual Decoding
May 22, 2026Jiahe Meng, Weiming Zeng, Yueyang Li +5Cross-Modal Knowledge Distillation Representation Learning
Lab of Digital Image and Intelligent Computation, Shanghai Maritime University, Shanghai 201306, China · Department of Language Science and Technology, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong SAR, China · Department of Neurology, Affiliated Lianyungang Hospital of Xuzhou Medical University, Lianyungang 222002, China +1
Brain-to-Image Retrieval and Reconstruction via Multimodal EEG Alignment
May 18, 2026Chi Kit Wong, Yan Liu, Haowen YanEEG-Based Visual Decoding Cross-Modal Alignment

May 23, 2026J/K move · Enter open · S save

MindAlign: Bridging EEG, Vision, and Language for Zero-Shot Visual Decoding

Zexuan Chen, Sichao Liu, Runhao Lu, Huichao Qi+3

KTH, SWeden · University of Cambridge, UK · EPFL, Switzerland+2

EEG-Based Visual Decoding Cross-Modal Alignment Multimodal Contrastive Learning

Open paper AskPDF arXiv

Visual decoding from brain signals is a key challenge at the intersection of computer vision and neuroscience, requiring methods that bridge neural representations and computational models of vision. We introduce a tri-modal contrastive framework for EEG-based visual decoding that aligns EEG, visual, and textual representations within a unified latent space. Our approach follows a two-stage design. First, we pre-train an EEG encoder via masked reconstruction on unlabeled trials, learning spatio-temporal regularities that transfer robustly to downstream tasks. Second, we jointly align EEG, image, and LLM-generated textual descriptions through contrastive learning, where text supervision acts as a semantic regularizer that injects linguistic structure into the shared space without overwhelming the primary EEG-image signal. The encoder integrates subject-specific adaptation, graph-attention over channels, and temporal-spatial convolutional embeddings. On the Things-EEG2 200-way zero-shot benchmark, our framework achieves 54.1% Top-1 and 83.4% Top-5 accuracy, substantially exceeding the strongest prior baseline (32.4% / 64.0%), with paired Wilcoxon tests confirming significance (p < 0.01) over all in-subject baselines. We validate generalization on Things-MEG. Analysis reveals that compact embedding geometries (CN-CLIP) outperform much larger backbones, and that decoding aligns with established neurophysiology of visual processing. This work is a critical step towards robust, semantically-grounded visual decoding from non-invasive temporal neural signals. The source code is publicly available in https://github.com/anon-eeg/eeg_image_decoding.

Abstract

EEG-to-Image Reconstruction

Explore similar work

MindAlign: Bridging EEG, Vision, and Language for Zero-Shot Visual Decoding

STAMBRIDGE: Spectral-Temporal Amplitude-aware Mid-Feature Bridge for EEG Visual Decoding

Brain-to-Image Retrieval and Reconstruction via Multimodal EEG Alignment