Abstract
The visual features associated with vivid mental images remain poorly understood. We collected more than 229,000 vividness judgments for the 73,000 images in the Natural Scenes Dataset in a large online experiment with 1,991 participants. We then trained lightweight MLP readouts on frozen vision-model representations to predict average imagery-vividness ratings from image content. The best individual model reached r = 0.376 [0.355, 0.396], approximately 95% of the estimated image-level reliability ceiling (r ≈ 0.395). A principal-component analysis of predictions from 14 models identified a shared image-property component that explained 72.3% of cross-model prediction variance and correlated with the measured target at r = 0.378. The best tested predictor of variation remaining after this component reached r = 0.077, approximately 39% of the corresponding residual-target ceiling (r ≈ 0.200). Position, distractor, low-level, caption-derived, coarse-category, memorability, and participant-level analyses constrain the interpretation of this performance. Average vividness ratings are strongly predictable from image content relative to the reliability of the measured image-level target. However, much of the predictable signal is shared across model families and appears to reflect general image properties, while reliable variation remains beyond this shared component. The results therefore establish a robust association between visual content and average vividness ratings, without implying that the models explain the subjective experience of mental imagery; the observer-specific processes that contribute to this experience remain open.
Where models and humans agree
Each point is one of the 7,300 test images. The horizontal axis is the shared component of 14 model predictions. The vertical axis is the human vividness rating. Dashed lines mark the 30th and 70th percentile of each axis. Hover a point to see that image.
Underpredicted
Images are partitioned by independent top and bottom 30% thresholds on human vividness and on the sign-aligned shared component (r = 0.378). Agreement cells are where that component matches the human rating. Disagreement cells are the underprediction of human-high indoor and structured scenes, and the overprediction of visually salient single-subject images.
Citation
@inproceedings{mahdiani2026modeling,
title={Modeling the Vividness of Imagined Natural Scenes Reveals a Model-Common Image-Level Component in Vision Models},
author={Mahdiani, Morteza and Landry, Catherine and van den Bosch, Jasper JF and Gosselin, Fr{\'e}d{\'e}ric and Charest, Ian},
booktitle={The Fortieth Annual Conference on Neural Information Processing Systems},
year={2026},
url={https://openreview.net/pdf?id=XB3iY2q9Vb}
}