EEG2GAN: Transformer-based EEG-to-Image Generation

Implementation of a Transformer encoder coupled with a Conditional GAN to reconstruct visual stimuli from raw EEG brain signals.

Research Report

Results

Quantitative comparison on the MindBigData ImageNet dataset (569 classes).

Method	IS ↑	EISC ↑	K-Means Acc	FID ↓
ThoughtViz (2017)	4.12	0.211	8.2%	312.4
LSTM Baseline	6.15	0.419	20.4%	141.4
EEG2GAN (Ours)	7.10	0.478	20.4%	128.9

Visualizations

Performance Metrics

Quantitative results including dataset statistics, confusion matrices, and ablation studies.

Main Metrics	Dataset Stats

Confusion Matrix	Ablations

Signal Analysis

Power spectral density tracking across different cognitive states and manifold interpolation proving learned smoothness.

Sample Generations

Representative grids across diverse ImageNet categories.

Architecture

graph LR
    A[Raw EEG] --> B(Filter)
    B --> C(Transformer)
    C --> D{Concat}
    E[Noise z] --> D
    D --> F[EEG-Conditioned GAN]
    F --> G[Image]

Loading

• Encoder: Transformer with 2 layers and 4 attention heads (compared against an LSTM baseline).
• GAN: Standard Conditional DCGAN with Hinge Loss and DiffAugment for stability.

References

Foundational GAN Research

• Generative Adversarial Nets | Ian J. Goodfellow et al. | arXiv:1406.2661
• Deep Convolutional Generative Adversarial Networks (DCGAN) | Alec Radford et al. | arXiv:1511.06434
• Conditional Generative Adversarial Nets | Mehdi Mirza & Simon Osindero | arXiv:1411.1784

EEG-to-Image & Signal Synthesis

• EEG2IMAGE: Image Reconstruction from EEG Brain Signals | Prajwal Singh et al. | arXiv:2302.10121
• EEG-GAN: Generative Adversarial Networks for EEG Signals | Kay Gregor Hartmann et al. | arXiv:1806.01875
• Generating Visual Stimuli from EEG using Transformer-Encoder and GAN | arXiv:2402.10115
• Guess What I Think: Streamlined EEG-to-Image with Diffusion | Eleonora Lopez et al. | arXiv:2410.02780

Reviews & Survey

• A Survey on Bridging EEG Signals and Generative AI | Shreya Shukla et al. | arXiv:2502.12048
• Interpretable EEG-to-Image Generation with Semantic Prompts | arXiv:2507.07157