The Dream2Image dataset (Bellec, 2025) released the first paired corpus of pre-awakening scalp EEG, verbatim dream reports, and AI-generated images approximating the reported dream (38 participants, 129 triplets), and its limitations and future-work sections flag what stands between this corpus and large-scale, generalizable decoding of dream imagery from EEG: the small sample size and the overfitting risk it creates for machine-learning and deep-learning models, the ceiling imposed by dream-recall quality and generative-model capability, biases inherent to AI image generators, and the need for continued experimental protocols with systematic dream-report collection. No published work has yet taken up any of these directions, so whether large-scale, generalizable decoding of dream imagery from pre-awakening scalp EEG can be achieved remains an open question.

Background

Electroencephalography (EEG) records the electrical activity of the brain through electrodes placed on the scalp; in sleep research it is used to identify sleep stages, including REM (rapid eye movement) sleep, the stage in which dreams are most frequent and most visually elaborate. A long-standing methodology collects dream reports by serial awakenings: participants are woken at targeted sleep stages and immediately report their preceding experience, yielding paired brain-signal/report samples. Two results anchor this paradigm: Horikawa, Tamaki, Miyawaki and Kamitani (2013) demonstrated with functional MRI (fMRI, a technique measuring brain activity via blood-oxygen changes) that a decoder — a model trained to predict which stimulus category a brain-activity pattern corresponds to — trained on waking perception could classify and track the contents of visual imagery experienced during sleep, establishing that dream reports reflect decodable brain activity rather than post-awakening confabulation; and Siclari et al. (2017) showed that dreaming is associated with a local 'posterior hot zone' of EEG activity that distinguishes dream experiences from dreamless sleep across stages. Scalp EEG is far cheaper and more portable than fMRI but carries much less spatial information, and the upstream recordings used here come from studies such as Zhang and Wamsley (2019) on EEG predictors of dreaming outside REM sleep and Konkoly et al. (2021) on real-time communication with lucid dreamers. In parallel, machine-learning research on EEG-to-image generation — exemplified by DreamDiffusion (Bai et al., 2023), which uses contrastive pretraining of an EEG encoder (learning representations by pulling paired EEG/image samples together in an embedding space) together with a latent diffusion model (a generative model that iteratively denoises random noise into an image) — has produced image reconstruction from EEG of awake participants viewing known stimuli; cross-subject generalization of such models remains poor without large training corpora. Dream2Image (Bellec, 2025) combined these threads: from the DREAM database (a public repository of overnight sleep-EEG studies) it harmonized 17-electrode pre-awakening EEG segments (T-15, T-30, T-60, up to T-120 seconds before awakening), verbatim dream reports with one-sentence condensations, and images generated from the reports by a pipeline of AI agents (semantic extraction, prompt construction, image generation, selection). The author states that the corpus is limited by its small sample relative to the needs of machine learning and deep learning, which raises the risk of overfitting (models memorizing the small training set rather than generalizing) and should be complemented with transfer learning (reusing representations pretrained on large corpora) and data augmentation; that the fidelity of the generated images depends strongly on the quality of dream recall and on the constraints of current generative models; that generated images may reflect biases inherent to AI models and should therefore be interpreted with caution; and that continued experimental protocols in sleep research including systematic dream-report collection are recommended, with dream recall improvable through preparation and training combined with relaxation, breathing, or meditation exercises.

Problem Statement

The Dream2Image dataset paper leaves open the question of whether large-scale, generalizable decoding of dream imagery from pre-awakening scalp EEG into images can be achieved, and its limitations and future-work sections identify the following directions and obstacles along the way: (1) the 129-sample / 38-participant scale limitation, which increases the risk of overfitting and restricts the generalizability of trained models, and which the author suggests addressing through larger-scale, collaborative datasets; (2) transfer learning and data augmentation, suggested by the author as techniques to compensate for the current scarcity of training data; (3) the ceiling that dream-recall quality — described in the paper as the main limitation — and the constraints of current generative models place on how faithfully generated images can match the dreamed experience; (4) biases inherent to AI image-generation models, which the author notes mean that generated images should be interpreted with caution; and (5) continued experimental protocols in sleep research that include systematic dream-report collection, with the paper noting that dream recall can be significantly improved with preparation and training. The operative question is whether large-scale, generalizable decoding of dream imagery from pre-awakening scalp EEG into images is achievable, and the required deliverable is a reasoned, evidence-based assessment of that question. The limitations and directions above are context for that assessment rather than a conjunctive set of deliverables: the assessment must state what would have to be demonstrated for the question to be answered in the affirmative, and must argue for or against feasibility under the field's standard conventions of held-out-subject and held-out-session evaluation.

References

  1. Bellec, Y. (2025). Dream2Image : An Open Multimodal EEG Dataset for Decoding and Visualizing Dreams with Artificial Intelligence. arXiv:2510.06252, DOI: 10.48550/arXiv.2510.06252. https://arxiv.org/abs/2510.06252
  2. Horikawa, T., Tamaki, M., Miyawaki, Y., & Kamitani, Y. (2013). Neural Decoding of Visual Imagery During Sleep. Science 340(6132), 639-642, DOI: 10.1126/science.1234330. https://doi.org/10.1126/science.1234330
  3. Siclari, F., Baird, B., Perogamvros, L., Bernardi, G., LaRocque, J. J., Riedner, B., Boly, M., Postle, B. R., & Tononi, G. (2017). The neural correlates of dreaming. Nature Neuroscience 20(6), 872-878, DOI: 10.1038/nn.4545. https://doi.org/10.1038/nn.4545
  4. Konkoly, K. R., Appel, K., Chabani, E., Mangiaruga, A., Gott, J., Mallett, R., Caughran, B., Witkowski, S., Whitmore, N. W., Mazurek, C. Y., Berent, J. B., Weber, F. D., Türker, B., Leu-Semenescu, S., Maranci, J.-B., Pipa, G., Arnulf, I., Oudiette, D., Dresler, M., & Paller, K. A. (2021). Real-time dialogue between experimenters and dreamers during REM sleep. Current Biology 31(7), 1417-1427.e6, DOI: 10.1016/j.cub.2021.01.026. https://doi.org/10.1016/j.cub.2021.01.026
  5. Zhang, J., & Wamsley, E. J. (2019). EEG predictors of dreaming outside of REM sleep. Psychophysiology 56(7), e13368, DOI: 10.1111/psyp.13368. https://doi.org/10.1111/psyp.13368
  6. Bai, Y., Wang, X., Cao, Y., Ge, Y., Yuan, C., & Shan, Y. (2023). DreamDiffusion: Generating High-Quality Images from Brain EEG Signals. arXiv:2306.16934, DOI: 10.48550/arXiv.2306.16934. https://arxiv.org/abs/2306.16934