A Survey on Bridging EEG Signals and Generative AI: From Image and Text to Beyond

In plain words

Electroencephalography (EEG) records the brain's electrical activity through electrodes on the scalp. The signal is noisy and blurry in space, but it changes quickly in time. Researchers have started to train generative AI models to turn these recordings into pictures, sentences or sounds: a person looks at a picture, reads a word or hears a sound, and a model tries to produce something that matches what they perceived.

This survey gathers that work and sorts it into three groups: images, text and audio or speech. For each group it explains the brain activity behind the signal, the public datasets, the kinds of generative models used, how the signal is turned into features, and how results are scored. Its main message is that the field is still small and hard to compare. Datasets are few and varied, models generalize poorly to new people, and there is no shared benchmark, so it is hard to tell whether a reported gain comes from the model or from the setup.

TL;DR

• This is a survey. It trains no model and runs no new experiment; it compiles published work from 2017 to 2025 §3.
• EEG-to-image work mostly uses encoder-decoder models with GAN, VAE or latent diffusion decoders, often conditioned on EEG embeddings §4.5.
• EEG-to-text work mostly uses open-vocabulary sequence-to-sequence models, often a pre-trained BART with a transformer encoder on EEG §5.3 §5.5.
• EEG-to-audio/speech work usually maps EEG to a mel-spectrogram and renders audio with a neural vocoder such as HiFi-GAN §6.3 §6.5.
• Do not compare the survey's numbers across papers. The only like-for-like table is ZuCo, and even that mixes different runs §5.7 §8.

Problem

Non-invasive brain-computer interfaces are mostly confined to research prototypes, partly because EEG has low spatial resolution and a low signal-to-noise ratio §1 §8. Earlier surveys cover AI in EEG brain-computer interfaces broadly, or cover EEG-to-output research in one or two modalities §2. The survey states the gap it fills as a single comparison of EEG-based generative media across images, text and audio §2.

The survey names these open problems: small and heterogeneous datasets, limited cross-subject generalization, weak use of EEG-domain features, lack of interpretability, lack of standardized splits and metrics, and evaluation protocols that can overstate performance §8 §5.8.

Method

The survey follows a structured literature review §3:

1. Search. Google Scholar, PubMed, ScienceDirect, IEEE Xplore, the ACM Digital Library and arXiv were searched with terms that combine EEG or electroencephalography with text generation, image reconstruction, speech synthesis and generative AI. The terms were refined iteratively §3 fig. 3.
2. Window and sources. Work published from 2017 to 2025 was prioritized. Both peer-reviewed papers and influential preprints were included §3.
3. Inclusion. A study had to use non-invasive EEG, apply machine learning or generative AI, and produce text, images or audio §3.
4. Exclusion. Classification-only work (for example sentiment, motor imagery or emotion recognition) and work using fMRI, MEG or invasive recordings were excluded §3.
5. Extraction. For each study the survey recorded the stimulus modality, use case, addressed challenges, architecture (GAN, VAE, transformer, diffusion), EEG encoding, datasets and metrics, in a table §3.
6. Synthesis. Studies were grouped into image, text and audio, and compared within each group on use case, architecture, encoding and evaluation. A cross-modal comparison follows §3 §7.

The number of records screened and the number of included studies are not stated §3.

The survey gives background on the neural basis of each modality §4.1 §5.1 §6.1 and on the generative building blocks §4.3 §5.3 §6.3:

• EEG bands used in the text: delta 0.5–4 Hz, theta 4–7 Hz, alpha 8–12 Hz, beta 13–30 Hz, gamma 30–100 Hz §1.
• Encoder-decoder: an encoder maps the input to a compact latent code and a decoder rebuilds the target from it §4.3.
• VAE: the latent is a distribution. Training minimizes a reconstruction loss plus a KL divergence term that pulls the latent toward a standard normal §4.3.
• GAN: a generator makes samples from noise and a discriminator learns to tell them from real data; both are trained together §4.3.
• Diffusion: a forward process adds noise step by step; a learned reverse process removes it to generate. Latent diffusion runs this in the latent space of a VAE-like encoder §4.3.
• Text generation: autoregressive language models predict the next token given the context §5.3. The survey gives a typical EEG-to-text framework: a multi-layer transformer pre-encoder on EEG features, followed by a pre-trained BART encoder-decoder that generates text §5.3.
• Audio: signals are often converted to mel-spectrograms before modeling. HiFi-GAN uses multi-scale and multi-period discriminators §6.3.

The survey's own synthesis of a generic pipeline is:

# Cross-modal EEG generation, as the survey summarizes the literature  [§7]
# Input:  eeg recorded while a stimulus was shown  [§1]
# Output: generated image, text or audio  [§1]

def generate_from_eeg(eeg, subject_id):
    # Preprocess: filter, remove artifacts, extract features  [fig. 2] [§8]
    features = preprocess(eeg)
    # Spatial-temporal encoder: CNN, RNN/LSTM/GRU, hybrid or transformer  [§7]
    latent = spatio_temporal_encoder(features)
    # Subject-adaptive layer, since EEG patterns differ between people  [§7]
    latent = subject_adapter(latent, subject_id)
    # Align with the target space, often with contrastive loss  [§4.6] [§7]
    latent = align_to_target_space(latent)
    # Image: GAN, VAE or latent diffusion decoder  [§4.5]
    # Text: sequence-to-sequence language model  [§5.5]
    # Audio: mel-spectrogram decoder with a vocoder  [§6.5]
    return modality_generator(latent)

This is a synthesis across papers, not one published model. The survey does not give a single set of hyperparameters for it.

The one numbered equation in the survey is the sequence-to-sequence factorization for EEG-to-text. Given EEG signal \mathcal{E} and target sentence \mathcal{S} of length T over vocabulary \mathcal{V}, the probability of the sentence is the product, over positions t, of the probability of token s_t given the EEG and the previous tokens eq. 1 §5.5:

p(\mathcal{S}|\mathcal{E}) = \prod_{t=1}^{T} p(s_t \in \mathcal{V}|\mathcal{E}, s_{<t}) eq. 1

Hyperparameters

The survey trains nothing, so the table lists its review settings and the parameters of the datasets it summarizes.

Datasets

The survey's dataset table, reproduced with its reference keys. Columns are as in the paper tab. 1. The reference key is the bibliography entry in the survey.

Results

The survey reports no new results. The only table that compares methods on one benchmark is the ZuCo open-vocabulary EEG-to-text comparison. Setup: ZuCo with the 80/10/10 sentence split §5.7; the reference point is the open-vocabulary model of Wang and Ji §5.7. Values are percentages, compiled by the survey from each cited paper, and author-reported tab. 4.

Sources for the rows are the cited papers in tab. 4. Dashes mean the metric was not reported there. In the paper, BELT-2 + LLM (T5) has the best BLEU-1 to BLEU-3 and all three ROUGE-1 values, and ETS (BART) has the best BLEU-4 tab. 4.

Other headline claims, without numbers in the survey:

• EEG-to-image, ENIGMA (kneeland2026enigma): the survey reports a multi-subject model needing less than 1% of the trainable parameters of earlier subject-specific approaches, a 165× reduction in total parameters for 30 subjects compared with a subject-specific model, and adaptation to a new subject with about 15 minutes of calibration data §4.4.
• EEG-to-image, low-density EEG: a portable 8-channel system was reported to support image classification and reconstruction §4.4 §8.
• EEG-to-image, Alljoined-1.6M: the survey reports roughly log-linear gains in decoding as training data grows, on a 32-channel consumer system costing about $2,200 §8.
• EEG-to-audio/speech: the survey reports no numeric results. Its tables list metrics only tab. 5.

Metrics used across the surveyed work, as listed by the survey:

• EEG-to-image tab. 2: top-k accuracy; Inception Score, FID, KID, LPIPS and a diversity score; SSIM, PixCorr, CLIP score and feature distance (AlexNet top-k, SwAV); visual inspection or user studies.
• EEG-to-text tab. 3: BLEU, ROUGE, METEOR; BERTScore, BLEURT; WER, TER.
• EEG-to-audio/speech tab. 5: RMSE, PCC, MCD, SSIM and PSNR on spectrograms; WER, CER, BERTScore; FAD, MOS; Hits@k.
• Cross-modal comparison tab. 7: semantic fidelity is partly shared across modalities, perceptual quality and evaluation protocols are not, and the survey recommends common protocols.

Paper vs code

No official code.

The survey links three repositories in its appendix app. A: perceivelab/eeg_visual_classification (EEG-to-image baseline), MikeWangWZHL/EEG-To-Text (EEG-to-text baseline) and youngeun1209/NeuroTalk (EEG-to-speech baseline). They are third-party projects and the survey does not show that the authors of this survey maintain them, so they are not used as this card's code.

Searches made: (1) the paper's text and appendix for code links; (2) GitHub repository search through the API for "EEG generative survey" and for the title words with the author names ("Bridging EEG Generative AI survey Shukla Mishra"). Both returned zero repositories.

Pitfalls

• A survey is not a specification. Nothing in this card can be implemented as one model. The pseudocode is the survey's generic synthesis §7. Re-implementing any single method needs the original paper.
• Numbers come from different papers. Each row of the ZuCo table comes from a different run, with its own model, preprocessing and training choices. The survey itself says that cross-study comparison is hard §8 tab. 4.
• Text scores can be inflated. Many open-vocabulary EEG-to-text studies use teacher-forced evaluation and lack control baselines such as random input, so BLEU and ROUGE may overstate how much the EEG contributes §5.8. The survey recommends noise-baseline testing, non-teacher-forced evaluation, significance tests with confidence intervals and effect sizes, and sentence-level splits with no overlap §5.8.
• Subject leakage inflates results. Random splits that place recordings from the same person in train and test sets overstate generalization. The survey reports that some benchmark datasets lose more than 30 F1 points under subject-independent evaluation, citing wang2026causesperformancedegradationcrosssubject §8. Use subject-disjoint splits.
• Dataset licenses differ from this card's license. This card is CC BY 4.0, which covers the survey's text and not the datasets it lists. Several datasets the survey lists are not free for commercial use: Alljoined-1.6M is CC BY-NC-SA 4.0 tab. 1. Some are "on request", "unavailable" or "not publicly available or specified" tab. 1. Several have no stated license: ImageNet EEG, ThoughtViz and KARA ONE tab. 1. The survey does not state the terms for the ImageNet images used in the ImageNet EEG dataset §4.2. Check each dataset's license before training a commercial system on it.
• A citation in the ethics section looks wrong. The text describes California SB 1223 but cites the Colorado key colorado2024hb1058 §9.4. Check the source before quoting the California law.
• Figures are not reproduced here. The card does not carry any figure image. Figure numbers cited as fig. 1–fig. 3 follow the compiled order in the fetched LaTeX source.
• Version. The card reflects the fetched LaTeX source for arXiv v4 (updated 2026-09-15). Section and table numbers follow that source.

Lineage

• Builds on: none. The survey positions itself against earlier surveys, cited in §2 as cao2020review, sabharwal2024comprehensive, eldawlatly2024role, habashi2023generative, barbera2026using and wang2026generative. They are named only, and no content from them is used on this card.
• Compared against: the same related surveys §2.
• Superseded by: not stated.