Studying the qualities of conscious mental images is difficult because introspective reports are inherently subjective and offer limited, variable access to perceptual detail. The authors propose a neuro-computational strategy that substitutes introspection with an empirical, hypothesis-driven pipeline: generate candidate visual stimuli that approximate imagined content, simulate cortical responses to those stimuli, and test which stimulus properties best align with neural activity recorded during imagery. This approach aims to capture objective signatures of imagery content and quality without direct self-report.
To enable this approach the study collected a large-scale EEG dataset. Ten participants each completed ten sessions, yielding a total of 43,200 imagery trials. On each trial participants were asked to imagine one of 16 scenes prompted by text. The dataset was designed to sample imagery-related EEG activity across repeated, controlled prompts, providing a basis for comparing neural signals to simulated responses evoked by candidate images.
The authors used AI image generation to create sets of candidate images corresponding to the imagined text prompts. These AI-generated images served as concrete approximations of the putative visual content participants might hold in mind. For each candidate image, the team computationally simulated responses of the visual cortex. The simulation pipeline and the specifics of the models used are reported in the source code and data repositories referenced by the authors (see Footnotes and Data/Code links). This two-step procedure—AI-based candidate creation followed by neural-response simulation—provides the testable stimuli needed to probe representational alignment with recorded EEG.
The core analytic step was the assessment of representational alignment between simulated cortical responses to candidate images and the rhythmic EEG signals recorded during imagery. The authors focused on alignment with oscillatory activity rather than isolated time-domain patterns. In line with prior literature linking imagery and visual processing to rhythmic signatures, the analyses emphasize correspondence with EEG power and representational structure across frequencies, with particular attention to alpha-band activity.
Analyses revealed that mid- to high-level features derived from the AI-generated candidate images showed reliable representational alignment with human EEG during imagery, notably with alpha activity. This outcome is consistent with earlier reports implicating alpha-band dynamics in visual imagery and top-down visual processing. The finding indicates that features beyond low-level pixel structure—those reflecting more gestalt or semantic image aspects—are informative when approximating neural imagery representations.
To test how image qualities influence alignment, the authors systematically manipulated properties of the candidate images and reassessed representational correspondence with cortical imagery signals. Two manipulations increased alignment: spatial blurring and reduced contrast. These manipulations suggest that imagined visuals are often characterized by degraded sensory detail relative to veridical perception. The authors also report that an image style described as “psychedelic,” characterized by flowing textures and proportion distortions, yielded increased alignment, suggesting that mental images may incorporate global distortions or dynamical wrapping of form.
Taken together, the results support two complementary characterizations of mental imagery. First, imagery often exhibits reduced sensory quality—less spatial detail and lower contrast—relative to external images. Second, imagery may include more complex global transformations that distort proportions and impart flowing or emergent features (the authors describe this qualitatively as a psychedelic-like style). The neuro-computational pipeline thus provides objective evidence for these properties by linking manipulated image features to EEG signatures recorded during imagery.
The authors make data and code available via public repositories; footnotes in the source point to a GitHub repository, an OSF overview, and a Hugging Face dataset containing the image sets. The study is a preprint and has not been peer reviewed; the source explicitly reports that status. Specific details about the models used for image generation, the parameters of the neural-response simulation, and statistical effect sizes are available in the linked code and supplementary materials but are not exhaustively reported within the source abstract. The authors declare no competing interests.
This hypothesis-driven, non-introspective strategy complements reconstruction approaches to imagery by testing which candidate-image properties best explain neural imagery signals. The method may be extended to other stimulus classes, more diverse participant samples, or multimodal neural measures. Because the present work is a preprint, replication and peer-reviewed validation will be important next steps to confirm generalizability and to further specify which computational features most robustly map onto subjective imagery experiences.