The temporal and spatial properties of memorability reveal insights into the art creation process

Abstract

When creating art, artists must make constant decisions with each pen stroke in order to achieve their goals of the piece. One common goal is to create something memorable—a piece that will persist in the memories of the observer. However, an open question is how different artistic decisions impact the memorability of a piece as it is being created. To test this question, we utilized a neural network (ResMem) to track the frame-by-frame changes in memorability for 50 videos of the creation of digital art pieces. We tested this neural network’s predictions via a human memory experiment (N=399), finding that ResMem’s predictions significantly correlated with human memory performance across the art creation process. We then collected multiple measures of each piece’s creation process, including low-level visual features, mid-level visual features, neural network-derived features, artistic features, and spatial properties. We find that changes in details and low-level visual features do not contribute to the memorability of a piece; rather more global measures of simplicity and variability relate to the memorability of a piece. Further, larger changes (e.g., blocking, sketching), particularly below the horizon of the image, make the biggest impacts on memorability. Importantly, we find that the memorability of a piece is largely set from the very beginning, and changes that occur later have diminishing impact on the memorability of a piece. These findings have important implications for both artists and psychologists in thinking about the factors that impact memory during the creation of an image.

1. INTRODUCTION

The creation of artwork is a dynamic, complex process that involves the balancing of multiple goals of the artist. One of these goals is to create an impactful piece that lasts in the memory of the viewer, so that they engage with the piece even after it is out of sight. However, not all art that we see will last in memory, and our likelihood to remember a piece is surprisingly similar across people and predictable from the piece itself (Davis & Bainbridge, 2023).

Prior work has found that pieces of artwork have an intrinsic memorability (Davis & Bainbridge, 2023), meaning they can be assigned a “score” reflecting the likelihood they would be remembered (e.g., everyone would remember a 100% memorable piece) (Bainbridge, 2019). Specifically, using a dataset of over 4,000 paintings from the Art Institute of Chicago, artworks were shown to be consistently remembered or forgotten across people in both an online study and in a naturalistic in-person museum visit (Davis & Bainbridge, 2023). Importantly, ResMem, a deep learning neural network (DNN) trained on predicting memorability of real-world photographs in an experimental setting, could predict the art pieces people remembered both online and in the museum visit (Needell & Bainbridge, 2022). Further, an artwork’s level of fame as judged by the museum curators was correlated with its intrinsic memorability judged by ResMem. This suggests that memorability is a tractable property for pieces of art that directly impacts what a viewer will remember in a museum, and what will become famous more globally. Indeed, recent convergent work has found that memorable image posts on Reddit also tend to receive more engagement and comments (Peng & Bainbridge, 2024). Crucially, memorability is not easily reducible to an alternate visual property such as aesthetics, color, or edges (Isola et al., 2014). It is also separable from other cognitive phenomena—independent from attentional state (Wakeland-Hart et al., 2022), motivational state (Bainbridge, 2020), and preference (Isola et al., 2014). Image memorability is thus an isolable image property, currently thought to reflect higher order semantic properties of an image (Kramer et al., 2023) and how easy it is to process by the brain (Goetschalckx, 2019a). However, it is still largely unknown what specific properties make an image or piece of artwork memorable. Examining the impact of artistic decisions during the creation process on memorability may be able to help answer this question.

Many previous studies have utilized the artwork creation process—specifically drawing—as a tool to quantify participants’ memories (Wammes et al., 2016; Wammes et al., 2018) and other cognitive processes (Fan et al., 2023). Drawing has been cleverly utilized to uncover different properties of memory, such as its relationship with concreteness effects (Roberts & Wammes, 2021) and differences in encoding between object and spatial memory (Bainbridge et al., 2019; Megla et al., 2024), that are inaccessible using more traditional memory tasks. Drawing has even proven itself to be an effective way to improve one’s own memory (Wammes et al., 2016) and has provided fascinating insights into the differences between the memories of artist and non-artists, showing that drawing experts have better memory (Perdreau & Cavanagh, 2015; Děchtěrenko et al., 2024), insight into an image’s memorability (Davis et al., 2025), and visual abilities while drawing (Chamberlain et al., 2019) than novices.

Drawing as an experimental method is particularly exciting because it reveals the process of visual production (Fan et al., 2018); researchers can analyze the dynamic image creation process and the motivations behind creative decisions, in contrast with the static images and photographs more commonly used in visual memory and perception research. In other words, by seeing how artwork is created from scratch, we can understand how percepts and memories of an image are built element-by-element rather than having to disentangle the contributions and interactions of elements within a completed image. This “bottom-up” approach is especially valuable in a new era of visual perception and aesthetics science, where many computational visual features have emerged that are correlated with the subjective aesthetic experience of a piece, and that can be automatically quantified from an image, such as self-similarity, complexity, symmetry, entropy, object integration, and metabolic cost (Damiano et al., 2023; Nara & Kaiser, 2024; Redies et al., 2025; Tang et al., 2025). For example, Redies conducted a case study of five of his own abstract artworks during the creation process, and observed how self-similarity and sparseness evolved over the creation of these five pieces (Redies, 2020). These computational approaches thus enable one to easily quantify how visual feature measures change stroke-by-stroke as a piece is created, to create a real-time estimate of the perceptual and aesthetic experience.

Emerging work in neuroaesthetics has also begun to uncover interesting relationships between art and memory (Chatterjee & Vartanian, 2014). Several models of neuroaesthetics, including Leder and colleagues’ Model of Art (Leder et al., 2004) and Tinio’s Mirror Model of Art (Tinio, 2013), incorporate memory as a key component of our ability to have an aesthetic experience, as part of that experience requires relating what we see to what we know. The relationship between memory and art has also been tested in the setting where art is most often found, museums and galleries. For example, previous studies have shown that memory for artwork is greater in the museum than in the laboratory setting (Henkel, 2014), photographing whole objects impairs memory of their details (Brieber et al., 2015), and memories of artworks from a virtual museum tour are vulnerable to intrusions of false memory “lure” artworks (Daviddi et al., 2022). Therefore, the way in which we engage with artwork clearly impacts our memories, and memory also plays a central role in our understanding of art.

It is thus essential to understand how artistic decisions impact the future memory fate of a piece. Most notably, as drawings, artworks, logos, designs, or advertisements are being created, they are often analyzed or tested at their completion rather than throughout the creation process. However, it is possible that the memory fate of a piece is determined much earlier in the creation process, and understanding this key point may be important in helping artists create memorable pieces. The dynamic and goal-directed nature of art creation may give new insight into factors that impact human memory, given that we can observe how smaller-scale changes in an image impact memory performance. Meanwhile, understanding how visual and mnemonic features change as a piece of art is created can also inform our understanding of how we create and view aesthetic experiences.

In this study, we test how memorability changes over the art creation process, and examine how artistic, visual, and spatial factors impact the memorability of a piece. To answer this question, we collect videos of 50 digital artworks being created and track their memorability over time using the ResMem neural network. We validate ResMem’s predicted memorability scores through a human memory experiment, confirming that ResMem is an effective way to automatically predict memory for steps throughout the art creation process. Finally, by examining the visual, artistic, and spatial features of artworks at key points in the artwork creation process, we uncover temporal and spatial properties of memorability that have the potential to impact both memory researchers and artists alike.

2. METHODS

In this study, we collected 50 YouTube videos of a variety of digital artworks as they were created. These videos were then broken down into, on average, about 50 still images each to be used for data analysis. We used ResMem, a deep learning neural network (Needell & Bainbridge, 2022), to predict the memorability of the images to track each artwork’s memorability over time. We also ran a human memory experiment (N=399) to confirm ResMem’s predictions for three main points of the artworks’ creation: the start, composition, and final images. To better understand the impact of different changes throughout the creation process, two artists coders categorized the changes that occurred between subsequent images, as well as categorized final images based on their viewing perspectives. To determine the influence of visual features on memorability, we investigated the role of a set of low-level, mid-level, and neural network-based visual features on memorability. Specifically, for low-level features, we tested the role of each image’s hue, saturation, value, overall coloration, and spatial frequency on memorability. For mid-level features, we tested complexity, clutter, mirror symmetry, self-similarity, homogeneity, anisotropy, and sparseness. For computer-vision based features, we tested the relationships of image integration, metabolic cost, and distinctiveness with memorability. Lastly, we tested how the location and size of changes between images impacted memorability of a piece.

2.1. Stimuli

To test memorability during the art creation process, we collected a total of 50 videos demonstrating the creation of a piece of digital artwork from 23 different artists on YouTube. These videos generally show the full creation of the piece and are often viewed by other artists who are either fans of the artist or hope to learn from the video. Since these digital artworks can take many hours to complete, most are recorded in a timelapse video; this sped-up video type is especially common for digital artists to use given that popular drawing apps like Procreate (https://procreate.com/procreate) have built-in features that allow you easily record a timelapse video of your artwork from start to finish. To collect these 50 videos, we searched YouTube using terms like “digital drawing,” “timelapse,” and “tutorial.” Oftentimes, YouTube’s suggested videos were very helpful in finding relevant videos. While collecting videos of artworks for our dataset, we aimed to collect a balanced and unbiased sample, so that no single style or subject matter was over-represented. The completed artworks in the videos covered an array of artistic styles including hyper-realism, illustration, and cartoon. There was a wide range of subject matter including portraits, landscapes, figures, objects, and architecture, each within several artistic styles. While the final pieces cannot be shown here for copyright reasons, we provide links to the YouTube videos on our Open Science Framework repository (https://osf.io/9z5rj/).

In addition to artistic and stylistic considerations, many technical components were accounted for. We selected videos where only the artwork was displayed on the screen (no hands, drawing tools, cursors, etc.) and the time lapse had to be at a constant speed. The entire artwork had to be visible throughout nearly the entire video (e.g., there could be no zooming in to work on details), and the video had to play from the near beginning of the art-making process until the end. These videos also had a fixed “camera” (the screen-recorded view of the art creation program, usually Procreate), so that there was no translation or rotation of their view. We did not restrict the aspect ratio of the drawing; the drawings collected were a variety of portrait (N = 9), landscape (N = 37), as well as square (N = 4) orientations.

From each of these videos, we extracted still images at semi-regular intervals. Overall, the number of images taken was dependent on the length of the video and the complexity of the artwork. We collected the videos and still images in two waves (one of 30 videos, one of 20 videos). For the first 30 videos, the number of still images depended on the length of the video and the number of key changes throughout the video (e.g., filling in color, adding an object), whereby shorter videos and simpler artworks tended to result in fewer still images. Specifically, the first author (artistically trained, Bachelor’s in Visual Arts) selected frames based on when he detected a noticeable difference between frames within the overall artwork. However, we adopted a more standardized approach for the latter 20 videos. For those videos, we collected still images at regular time intervals such that about 50 frames were collected per video to roughly match the average number of still images from the first 30 videos.

We also excluded some of the still images for various reasons. The most common reason for exclusion was when images interrupted the continuity of the art-making process. Since these were digital artworks, artists could toggle layers of the artwork on and off and change the opacity of layers while working on another. We believed that this toggling of layers, creating such drastic and fast changes, was not actually reflective of changes being made to the artwork; rather, it is simply a tool for the artist to better understand parts of the image before moving forward, which is why these images were removed from the set of still images. Artists also sometimes mirror-flipped the image to ensure symmetry or consistency of their piece. Such flipped images, almost always across the Y-axis, were flipped back to their original orientation for their still images. Lastly, images that had an external object, like watermarks or logos in the beginning or end of the video, were excluded. In total, 2,538 still images were collected from the 50 videos (M number of images per video = 50.8, SD = 20.2). A histogram showing the number of still images per video and an example of the still images collected from the artworks videos is shown in Figure 1.

Figure 1: Histogram and Examples of Still Images Collected From Each Video.

(Left) A histogram of the number of still images collected for each of the 50 videos. (Right) Examples of still images collected from each video. The example images shown to the right are a depiction of what the still images look like (drawn by the first author); they do not come from the 50 videos in the stimulus set, due to copyright reasons.

Additionally, the still image at which the artwork composition was complete was marked for each video. We defined the composition image as the still image containing all of the main objects that are present in the final image in their appropriate location. It can exclude the smaller or less relevant objects, details, colors, highlights, and shadows, but contains the arrangement of the final image. These composition images are used as key markers during the art-making process in later analyses.

2.2. AI-based Quantification of Still Image Memorability

To obtain the memorability scores for the 2,538 still images collected from the 50 videos, we used ResMem, a publicly available deep learning neural network (DNN) designed to predict the intrinsic memorability of an image (Needell & Bainbridge, 2022). Each still image was inputted into ResMem, and it returned a predicted memorability score, ranging from 0 (completely forgettable) to 1 (completely memorable).

ResMem has been shown to generalize to a variety of image types and tasks due to its two-branch processing architecture (Li et al., 2022; Kramer et al., 2023; Guo & Bainbridge, 2024). Its first branch is an 8-layer shallow neural network capable of successfully predicting object categories with representations similar to those within the human visual processing stream (Cichy et al., 2016), and similar in architecture to AlexNet (Krizhevsky et al., 2012). The architecture of its second branch comes from ResNet-152, a 152-layer neural network designed to identify semantic information through its depth and ability to skip layers through residual connections (He et al., 2016). This deeper processing allows for ResMem’s improved predictions in comparison to previous networks, especially those with shallow architectures (Khosla et al., 2015). ResMem was trained and validated using a combined data set with nearly 70,000 images (Khosla et al., 2015; Goetschalckx et al., 2019b), mostly of photographs of objects and scenes, where the memorability of the images was estimated by their hit rate through a continuous recognition memory task. In terms of image pre-processing, ResMem pads all images with white pixels to make them square, and resizes them to 227 × 227 pixels before computing their memorability score.

Importantly, a previous study which used a dataset of 4,021 paintings from the Art Institute of Chicago showed that ResMem was able to successfully predict the memorability of artwork in both an online memory experiment (Spearman ρ= 0.45, p < 0.0001) and a naturalistic in-person memory test in a museum visit (Regression β = 0.18, p = 0.020) (Davis & Bainbridge, 2023). As a result of ResMem’s success in predicting the memorability of artworks, we chose to use it to predict the memorability of steps throughout the artwork creation process here.

2.3. Human-Based Quantification of Still Image Memorability

2.3.1. Motivation and Stimuli

Despite ResMem’s success in predicting the memorability of artwork in a previous study (Davis & Bainbridge, 2023), we validated these results by running a memory experiment to determine the memorability scores of each of the start, composition, and final images. If ResMem was able to successfully predict the human memorability scores for these images, it would validate our use of ResMem for the remaining images that fall in between these three major timepoints. One of the start images was excluded because it was extremely difficult to differentiate from just a white screen, meaning that 49 start images, 50 composition images, and 50 final images were included in the experiment.

2.3.2. Participants

A total of 399 adults participated from the online experiment platform Prolific (196 female, 197 male, 6 other/unknown; M_age = 36.3, SD = 11.4). Participants were prescreened such that they were between the ages of 18 and 64, fluent in English, and had no vision or color vision impairments after correction. Participants also had to correctly answer three basic English common-sense questions and correctly answer at least 50% of targets, foils, and attention checks in the task for their data to be included. The number of participants to recruit was determined as the number required to accumulate an average of at least 40 ratings on each image, as 40 ratings has been shown in prior work to be sufficient to estimate memorability scores for a single image (Bainbridge et al., 2013). Participants consented to participation and were compensated for their time. The full study protocol was approved by the university’s Institutional Review Board.

2.3.3. Task Procedure

Participants engaged in a continuous recognition memory task, as has been used in prior work to obtain memorability scores for photographs (Isola et al., 2011; Bainbridge et al., 2013). Participants viewed a stream of 76 images, each followed by a fixation cross, and had to press the “R” key whenever they identified a repeated image from earlier in the sequence. The start, composition, and final images were separated into 3 different memory experiments with different participant samples. Each image and fixation cross were presented for 1,200 ms, with an 800 ms interstimulus interval. 15 of these images were target images, for which their repetition occurred at least 30 s later. The remaining images were filler images, for which some repeated after a brief interval (1 to 5 images) to maintain participant vigilance. From the participant perspective, nothing differentiated targets from filler images.

For each image, we then calculated its hit rate (HR) as the proportion of participants who correctly identified that image as a repeat, amongst those who saw it as a repeated target. We also calculated its false alarm rate (FA) as the proportion of participants who pressed the “R” key on the first presentation of the image, amongst those who saw it as a target. We calculated the memorability score for each image as the corrected recognition (CR) rate, or HR–FA. Prior work has shown that memorability effects hold regardless of whether HR, FA, CR, or d-prime are used as the measure of memorability (Bainbridge et al., 2013). The experiment took approximately 5 min and participants were compensated $1.00 each for their participation.

2.3.4. Consistency Analysis

To test if participants showed agreement in the images that they remembered and forgot, we performed a split-half consistency analysis across 1,000 iterations. For each iteration, the participant pool was split into two random halves, and a CR score was calculated for each image within each participant half. We next calculated the Spearman rank correlation of the two sets of CR scores (from each participant half) and corrected it with Spearman–Brown split-half reliability correction. The final consistency is the mean correlation coefficient across the 1,000 iterations. A p-value was estimated with a permutation test, by shuffling the image order for one of the two participant halves, and then correlating it with the other half. This provides an estimate of the null hypothesis (no relationship between participant halves). The final consistency was compared against the distribution of these 1,000 shuffled null correlations to give a final p-value.

2.4. Image Categorization Task

In order to analyze decisions in the art creation process, two artist coders categorized the changes between still images within a video as belonging to one of three categories: sketching, blocking, and details. These coders had a Bachelor’s Degree in Visual Arts. Each coder labeled all images, and then for any labels on which they disagreed, they discussed and determined a final label by consensus. They were told to label these frame-by-frame changes based on the following categories (Table 1).

Table 1:

Instructions for artist coders for labeling artistic changes

CATEGORY	DEFINITION
1 – Sketching	- Linework that generally lays out the components of an image - Can either be over the entire image or just a subsection of the image - Usually done as a guide for other changes
2 – Blocking	- Filling in parts of the image - Building / filling in the different objects in the image - Generally on the scale of whole objects / multiple objects / large components of an object - Includes shadows / lights that span multiple objects or large sections of main objects - Also includes shifts in scene (ex: zoom in / out) to change the image’s composition
3 – Details	- Changes generally done within individual objects - Can include lines, shadows, and highlights on an object or smaller areas of multiple objects - Can also include smaller components that span the entire image (ex: thin streak of light)
Other Notes	- If changes occur that fit into more than one category, chose the category that had the greater change (ex: if there is a lot of sketching, and just a small amount of blocking, select “1”) - Please keep in mind that the change can be the addition or removal of one of the 3 categories. Also, changes from each category can be in black & white or color.

The same artist coders were also shown all of the final 50 images, and were instructed to categorize them into one of two categories based on whether they did, or did not, have a 1-point perspective: a technique that shows how objects appear smaller as they get further away, converging towards a single “vanishing point.” Artworks, and images in general, can have a variety of perspectives ranging from zero-point (no vanishing point), to 5-point (fisheye), and ultimately as many as the artist would like (multi-point). In this task, we only asked artist coders to categorize the artworks into two groups for two main reasons: first, this was only meant to serve as a preliminary investigation into the impact of perspective; second, the perspective of many artworks can sometimes be subjective and difficult for artists to agree upon, so we chose a simpler approach.

2.5. Quantifying Visual Features

2.5.1. Low-level Visual Features

Visual features including Hue, Saturation, Value, Spectral Energy, and other color features were determined using a MATLAB toolbox (Bainbridge & Oliva, 2015). First, all 2,538 images collected were converted from RGB color formatting into HSV color formatting, where each pixel is composed of three layers: hue, saturation, and value.

Hue.

Hue is measured from 0 to 1, but maps onto a color wheel that typically ranges from 0 to 360 degrees (where 0 = 0 degrees and 1 = 360 degrees), where 0 = Red, 60 = Yellow, 120 = Green, 180 = Cyan, 240 = Blue, and 300 = Magenta. Hue Mean is measured as the mean hue pixel value across an image, and Hue Contrast is measured as the standard deviation. While hue provides important information about an image, it often becomes uninterpretable because of its cyclical nature. For example, the average hue of an image that is primarily red (hue = 0) and magenta (hue = 300), will be about 150 (yellow-green), despite there being no yellow or green pixels. To correct this, we also categorized each image into 18 distinct color categories based on all three channels of the HSV color spectrum (see Supplementary Table 1).

Saturation.

Saturation is measured from 0 to 1, and represents the intensity of a color, or the color’s distance from white. A saturation of 0 means that the pixel is completely greyscale, whereas a saturation of 1 is the most intense possible color. Saturation Mean is measured as the mean saturation pixel value across the image, and Saturation Contrast is measured as the standard deviation.

Value.

Value is measured from 0 to 1, and represents how light or dark a color is, or the color’s distance from black. A value of 0 means that the pixel is completely black, and a value of 1 means that the pixel can range from being completely colored (ex: pure red) to completely white. Value Mean is measured as the mean value pixel value across the image, and Value Contrast is measured as the standard deviation.

Spectral Energy.

Spectral energy is a measure of spatial frequency and is dependent on the directions and distributions of the edges and lines in the image. Here, spectral energy was defined as the spatial frequency that contains 80% of the power spectrum energy of the image.

2.5.2. Mid-level Visual Features

We calculated several mid-level visual features for the artworks using the Aesthetics Toolbox (Redies et al., 2025), a publicly available tool that can calculate many low- and mid-level computer vision metrics shown to correlate with aesthetics ratings. We specifically tested five mid-level visual features: complexity, self-similarity, homogeneity, anisotropy, and sparseness. Briefly, complexity reflects the amount of changes across the pixels of an image, using the histogram of oriented gradients (HOG) method. Self-similarity measures the similarity of local parts of the image to the whole (e.g., fractals are highly self-similar). Homogeneity quantifies how scattered the elements are in an image (i.e., more scattered means less homogenous) and is related to perceptions of the balance of the image (Hübner & Fillinger, 2016). Relatedly, anisotropy tests the heterogeneity of gradient strength across an image, using HOG. Sparseness tests the variance of early-layer filter responses of the AlexNet neural network for an image, reflecting how heterogeneous the image is across local subregions. For more details about these measures, please refer to Redies et al., 2025.

We also computed two other mid-level features separately: clutter and global symmetry.

Clutter.

In order to measure clutter, images were first converted to greyscale and an “edge-image” was created by filtering the greyscale image with a Laplacian of Gaussian (LoG) filter through a MATLAB edge detection function, edgeim. The edge-image is only composed of white pixels (edges) and black pixels (non-edges), and Clutter is measured as the proportion of white pixels to total pixels. Clutter can range continuously from 0 to 1, where 0 means there are no edges in the image, and 1 means that the image is entirely made up of edges.

Global symmetry.

To calculate global symmetry, we used the approach by Damiano et al., 2023. Specifically, we iteratively correlated the grayscaled pixel values of an equal number of rows or columns around each horizontal or vertical axis. This correlation was then downweighted by the proportion of pixels utilized by the correlation. We then identified the highest weighted correlation (highest level of symmetry) along the rows (“horizontal symmetry”) and columns (“vertical symmetry”).

2.5.3. Neural Network-derived Visual Features

We also tested three additional visual feature measures derived from neural network measures that have been recently attributed to ratings of aesthetics: image integration, metabolic cost, and distinctiveness.

Image integration.

This measure reflects the degree to which separate components of an image relate to the representation of the whole image. To calculate image integration, we followed the approach by Nara & Kaiser, 2024. Specifically, we extracted the layer outputs of the VGG-16 DNN for each image, as well as two half-versions of the image (non-overlapping 8×8 checkerboard versions of the images overlaid with gray tiles). We focused specifically on the outputs of convolutional layer 11, which showed the highest correlations with beauty ratings in Nara & Kaiser, 2024. To calculate image integration, we averaged the layer outputs for the two half-images, and then correlated that vector with the layer outputs for the whole image.

Metabolic cost.

This measure reflects the amount of artificial neuronal activity necessary to process an image. To calculate metabolic cost, we used the approach by Tang et al., 2025, where we summed the positive artificial neuron activations for all the convolutional and fully connected layers from the VGG-19 DNN.

Distinctiveness.

This measures how similar an artwork is to other artworks, as well as how similar it is to itself at different creation stages (see Kramer et al., 2023). To measure distinctiveness, we extracted layer outputs from an image from the VGG-16 DNN, and then correlated them between images. We focused on two layers–an early layer thought to reflect low-level visual features (convolutional layer 2), and a late layer thought to reflect higher-level semantic features (the last convolutional layer). To measure distinctiveness across artworks, we took the average Pearson correlation of a piece’s layer outputs with those of all other pieces. A higher average correlation indicates lower distinctiveness.

2.6. Spatial Data

2.6.1. Difference Image Collection

To determine the spatial properties that impact an image’s memorability, we focused on the changes that occurred between images. By subtracting subsequent images from one another (Ex: 5 – 4, 4 – 3, etc.) we were able to track individual changes that led to the next image. This resulting “Difference Image” was used to track exactly where changes occurred.

However, these Difference Images could not be collected for all 50 artworks. Since many of the videos had black borders, they were automatically cropped in MATLAB to contain only the artwork. For some of the video frames collected, we were unable to create a difference image due to ambiguous or unstable borders. If we could not produce a difference image for all frames in a video, we excluded the entire video from spatial analyses. This was done to avoid over-sampling from specific portions of the creation process; for example, if several videos contained unstable borders toward the beginning of the art-making process, excluding only those images would mean that images later in the art-making process would be inappropriately over-represented. Ultimately, 31 out of the 50 videos were used for these analyses, resulting in a total of 1,556 difference images.

We created these Difference Images so that they show the changed pixels exactly as they appear in the subsequent image. For example, if a green mark were added on a red background, the pixels in the Difference Image would be green. A limitation of the Difference Image is that, due to how image subtraction occurs in MATLAB, they cannot show the addition of pure black pixels because the pixels in the difference images where no difference occurred were also pure black, meaning that there is no way to distinguish between the two. As a note, this approach also cannot tell us whether a difference involves an addition or a subtraction of an element between frames.

2.6.2. Image Processing for Spatial Data Collection

From these Difference Images, we then localized where artistic changes tended to occur across the creation of the piece. To account for the fact that the artworks had a variety of aspect ratios, the Difference Images were all compressed to a 1080×1080 pixel square. Next, they were each converted to a binary mask, which isolates the individual, continuous clusters of pixels in the difference image, represented by 1 or 0. Groupings of pixels smaller than 50 were removed to reduce noise. The resulting clusters reflect where the changes occurred between frames, and represent a continuous area of (non-black) pixels, that can comprise multiple objects or a part of a single object.

Once these clusters are determined, the center of mass (CoM) of each of the clusters was computed; this provided a single location for each of the clusters of pixels within a difference image. In other words, they reveal the location of the main changes between each pair of still images. Importantly, the CoM of each cluster does not necessarily lie within the cluster itself. For example, if a hollow circle is drawn, the CoM will lie at the circle’s center, which is not one of the pixels that was changed. Lastly, the mean location of all centroids in each Difference Image was calculated and recorded as the “center point.”

3. RESULTS

3.1. ResMem Predictions Correlate with Human Memory

First, we conducted a memory experiment to obtain human memorability scores for each of the 50 start, composition, and final images, to test ResMem’s predictions of human memory. The start images had the lowest average memorability scores (M = 0.78, SD = 0.13, min = 0.48, max = 1.00), the final images had generally high memorability scores (M = 0.81, SD = 0.11, min = 0.57, max = 0.976), and the composition images had the greatest average memorability scores (M = 0.82, SD = 0.11, min = 0.52, max = 1.00). A split-half consistency analysis across 1,000 iterations showed that participants had significant agreement in the artworks that they remembered and forgot (Spearman–Brown corrected ρ = 0.56, p < 0.001). This suggests that memorability scores are consistent across people for these three types of images.

We then tested whether the memorability of an image remains relatively the same across the stages of the art creation process (Figure 2). Indeed, we found significant Spearman rank correlations in human memory performance across the start, composition, and final images of the pieces (all ρ > 0.35, p < 0.01). We found similar results when looking within ResMem’s predicted memorability scores. Its predictions were correlated between the final and composition stages (ρ = 0.57, p < 0.001) and the final and start stages (ρ = 0.34, p = 0.017), although not between the start and composition stages (ρ = 0.26, p > 0.05). Overall, these results show that the memorability ranking of a piece remains relatively the same across the creation process–its fate is largely sealed from the outset.

Figure 2: Human and ResMem Memorability Scores.

A correlation matrix showing the Spearman rank correlations between human and ResMem memorability scores for the three stages of the artwork completion process: start, composition, and final images. Human memorability scores were measured by corrected recognition (CR) rate. Non-significant correlations (p>0.05) are in light gray.

Next, we tested the relationships between human memorability scores and ResMem’s predicted memorability scores using Spearman rank correlations (Figure 2). ResMem significantly predicted the memorability of the final piece, with large effect size (ρ = 0.62, p < 0.001). It also significantly predicted the memorability of the composition stage, with medium effect size (ρ = 0.35, p < 0.02). Note that these correlations are close to (or at) ceiling performance given that random halves of participants only show a correlation of ρ = 0.56 with each other. However, interestingly, while ResMem’s prediction for the start image could not predict human memory performance for the start image (p > 0.05), its predictions at the composition and final image could (both ρ > 0.28, p < 0.05). These results suggest that ResMem can predict the memorability of a piece throughout the creation process, except at the beginning when the image is relatively sparse.

3.2. How Does Memorability Change with the Creation of a Piece?

Our next question was how the memorability of the artworks changed over time. Across all 2538 images, the mean and spread of the ResMem predicted memorability score (M = 0.75, SD = 0.11, min = 0.42, max = 0.95) was extremely similar to ResMem predictions from a previous study which studied the memorability of 4,021 paintings from the Art Institute of Chicago (M = 0.73, SD = 0.12, min = 0.41, max = 0.98), where there was a significant correlation with online human memory performance (Spearman’s ρ = 0.45, p < 0.0001) (Davis & Bainbridge, 2023). Refer to Supplementary Figure 1 for a depiction of the trajectory of memorability scores for all 50 pieces.

We tested how memorability changed between individual images by taking the difference in memorability between subsequent images, resulting in a total of 2,488 “Difference Scores.” Since it is not possible to get the Difference Score for the first image from each video, there are 50 fewer Difference Scores than there are still images. A simple linear model was used to test if the absolute value of the Difference Scores could be predicted by image completion (%). The overall regression was significant (Adjusted R² = 0.059, F(1,2486) = 157.9, p < 0.0001), where image completion significantly predicted the absolute value of the Difference Scores (β = −2.41e-4, p < 0.0001). This significant relationship remains when a random effect of the painting is added to the linear model (Adjusted R² = 0.177; β = −2.44e-4, p < 0.0001). Further, we observe a better model fit if completion time is modeled with a logarithmic function (Adjusted R² = 0.188, β = −0.009, p < 0.0001), suggesting that later times have less impact on memory performance. However, if we predict the raw Difference Score (instead of the absolute value), no significant model nor predictor emerges (p > 0.05). This suggests that changes to the artworks later in the artwork creation process were less impactful on memorability than earlier changes, but not necessarily in a specific direction. This is further supported by the fact that when the absolute value of Difference Scores were split up by each half of completion (0 – 50%, 50 – 100%), first half scores were significantly greater than second half scores (t(1938.2) = 10.72, p < 0.0001).

We were able to observe broader trends in how memorability changed over time by testing the three distinct stages of completion for each painting (start, composition, and final). Overall, the average ResMem-predicted memorability tended to decrease: while the memorability of start images was not significantly different from the composition images (t(48) = 0.79, p = 0.43), it was significantly greater than the final images (t(49) = 3.16, p < 0.01), and composition images were also significantly greater than the final images (t(49) = 2.69, p < 0.01). In other words, pieces were judged to be less memorable as their details were filled in. Additionally, variance in memorability increased over time, where the variance for final images was significantly greater than start images (F(49,49) = 0.55, p < 0.05). This could be partially attributed to pieces becoming more distinctive from start to finish. Composition images were indeed significantly more visually distinctive than start images, as judged by an early layer of VGG-16 (t(48) = 9.69, p < 0.001) as well as a late layer of VGG-16 (t(48) = 8.64, p < 0.001). However, distinctiveness did not change from the composition stage to the final stage (both p > 0.05). We also tested whether composition images were more visually similar to their corresponding final images than their start images. However, we found no difference in similarity between start and composition images, and composition and final images for either layer of VGG-16 (both p > 0.05).

3.3. Details and Perspective Have Minimal Impact on Memorability

In the Image Categorization task, a pair of artist coders labeled the changes between subsequent images by categorizing them into one of three groups: sketching, blocking, and details (Figure 3). In total, the artists found that there were 284 sketching (11.4%, M=5.68 per piece), 1265 blocking (50.9%, M=25.30), and 935 detail (37.6%, M=18.70) changes across all 50 artworks. A vast majority of the artworks had at least one of these changes, with only 7 artworks skipping over the sketching stage and moving directly to blocking and details.

Figure 3: Change in Artwork Memorability by Change Type.

The absolute value of the change in memorability, as grouped by the type of change. Changes in details had significantly less impact on memorability than blocking and sketching changes. For both plots, error bars represent standard error of the mean.

While there was no significant relationship between the change type and the raw change in memorability, a 1-way ANOVA showed that both blocking and details, on average, produced greater changes in the absolute value of the change in memorability (F(2, 2481) = 96.6, p < 0.0001). This difference is especially apparent when looking at the mean absolute change in memorability for each group (Figure 2), as the impact of detail ($\stackrel{-}{\mathrm{x}}=0.012$) was significantly less than sketching ($\stackrel{-}{\mathrm{x}}=0.026$; t(329.89) = 7.81, p < 0.0001) and blocking ($\stackrel{-}{\mathrm{x}}=0.028$; t(1824.2) = 15.22, p < 0.0001), while the difference between sketching and blocking was not significant (t(482.45) = −1.33, p = 0.18). In other words, detailed changes had the smallest impact on memorability of a piece.

As for the Perspective Identification task, the artist participants found that there were 20 artworks that had a 1-point perspective, and 30 that did not. There was no difference in memorability between the two groups (t(42.2) = −1.68, p = 0.0995), although the 1-point perspective pieces had a qualitatively lower average memorability score ($\stackrel{-}{\mathrm{x}}=0.70$) than the ones without ($\stackrel{-}{\mathrm{x}}=0.75$).

3.4. Memorability Reflects Simplicity

We tested correlations of human memorability scores for the final images with several low-level, mid-level, and high-level visual features.

For low-level properties related to color (hue mean, hue contrast, saturation mean, saturation contrast, value mean, value contrast) and spectral energy, we observed no significant correlations with memorability (all p > 0.05). This mirrors prior work that also does not observe relationships between low-level visual features and memorability (e.g., Isola et al., 2013; Davis & Bainbridge, 2023).

For mid-level properties, we found no significant correlation between memorability and horizontal or vertical symmetry, complexity, and homogeneity. However, we did observe some interesting significant relationships as well. Memorable pieces were associated with more anisotropy (ρ = 0.42, p < 0.005), higher sparseness (ρ = 0.52, p < 0.001), lower self-similarity (ρ = −0.46, p < 0.001), and lower clutter (ρ = −0.36, p = 0.010). All of these visual feature measures were also significantly correlated with each other (all ρ > −/+ 0.50, p < 0.001), thus likely reflecting a singular construct. In combination, it appears that memorable pieces are generally simpler and more varied across the image. This supports the finding from the artist coders that adding more details to an image decreases its memorability.

Finally, for the neural network-derived features, we found no relationship between memorability and metabolic cost (p > 0.05), image integration (p > 0.05), or distinctiveness in comparison to the other pieces, regardless of neural network layer used for the analysis (p > 0.05).

3.5. Central Changes Have the Greatest Impact on Memorability

We next tested whether the locations of artistic changes predicted memorability. We found that there was a significant positive relationship between the number of pixels that were changed and the absolute difference in memorability (r(1558) = 0.30, p < 0.0001), meaning that as more pixels in the image change, the greater the memorability will change. Additionally, there was a significant negative relationship between the number of pixels changed and the raw difference in memorability (r(1558) = −0.11, p < 0.0001). In other words, the greater the change, the larger the loss of memorability. However, this negative correlation could be due to the fact that 31 out of the 50 artworks had a lower memorability for their final images than their start images.

We created Difference Images showing where changes occur from frame to frame of each art creation video (see Methods). Figure 4 shows a heatmap of the changes across all 50 artworks, and the Centers of Mass (CoM) of the changes by completion level of the image. Qualitatively, we can observe the greatest intensity of change close to the center of the image. There are several smaller “hotspots” that span the x-axis, however intensity along the y-axis is largely towards the center. This is not surprising, given that the bottom of an image is typically the floor or ground where there are fewer objects than the middle of the image, with even fewer objects in the sky or ceiling at the top of the image. We also can observe that there is no clear pattern of location of change over the completion of the image (see also Supplementary Figure 2). Overall, most changes tended to occur centrally, and spanning the horizon of the image.

Figure 4: Distribution of all changes across the 50 collected artworks.

a) This heatmap depicts the changes that took place across all of the still images of the 50 videos of the artworks being created. Regions shown in bright yellow represent areas of higher concentration of changes, whereas regions in dark blue represent areas of the least concentration of changes. It is important to note that this image was created based on the still images of the artworks, meaning that artworks with a greater number of still images are represented more than those with fewer. However, this figure illustrates the general areas in an image where the greatest density of changes occurred. b) This scatterplot shows the X- and Y-coordinates of the Centers of Mass of changes across all of the still images. The color of the dot indicates the completion time at which the change is made, in reference to proportion completion of the piece. Generally, there was no relationship between time of completion in the artwork and location of changes (see also Supplementary Figure 2).

We then analyzed the center of mass (CoM) and cluster locations within the difference images, and related them to memorability. The Difference Images had, on average, more than 30 clusters ($\stackrel{-}{\mathrm{x}}=32.4$, SD = 43.0) which together spanned about a quarter of the 1080×1080 px standardized image (mean range_x = 531.0 px, mean range_y = 500.0 px). Out of the 1556 Difference images, Outlier images (N = 12), all of which had 365 or more clusters ($\stackrel{-}{\mathrm{x}}=815.7$, SD = 379.3) were excluded; these images had an excessive number of clusters not because of the pixels that were added or erased in the image, but because the artist manipulated the entire image itself, often shifting all of the pixels in the image. This includes zooming in (N = 6), zooming out (N = 1), shifting the image (N = 2), compressing the image in the x or y axis (N = 2), and adding a texture across the image (N = 2); these do not sum to the 12 outliers because these effects are not mutually exclusive. Additionally, all images that had no clusters from filtering out noise were excluded (N = 120). After excluding outliers and images with no CoMs, we were left with 1,424 Difference Images with CoMs.

Using the 1,424 Difference Images, we looked to see if there was any relationship between the number and properties of CoMs and memorability by using a multiple regression model which included the number of CoMs, mean CoM location along each axis (X mean and Y mean), and the standard deviation of CoM location along each axis (X STD and Y STD) to predict the memorability score. We found that these properties could not predict changes in raw memorability (Adjusted R² = −0.00265, F(1,1418) = 0.247, p = 0.94). However, they could predict the absolute change in memorability between frames (Adjusted R² = 0.072, F(1,1418) = 23.1, p < 0.0001). Several variables including Y mean (β = 0.083, p < 0.01), X STD (β = 0.075, p < 0.05), and Y STD (β = 0.22, p < 0.0001) significantly predicted the absolute value of the changes in memorability. However, the number of CoMs (β = 0.018, p = 0.53) and the X mean (β = −0.045, p = 0.080) did not show a significant effect. This means that more spread-out changes that are lower in the y-axis produce greater changes in memorability. It is not surprising that more spread-out changes correlated significantly with absolute change in memorability, especially given our earlier finding that there is a significant relationship between the number of pixels that changed the overall change in memorability. However, it is uncertain why there is a bias toward changes lower in the y-axis.

To investigate the relationship between center point location and absolute change in memorability, we segmented the 1080×1080 px square images into five concentric “rings” (Figure 5) of equal width, ranging from Ring 1 on the outside of the image and Ring 5 at the center of the image. The number of change center points per ring were as follows: Ring 1 = 48, Ring 2 = 219, Ring 3 = 415, Ring 4 = 444, Ring 5 = 298. We found that the mean absolute change in memorability decreased as the ring number increased, meaning that center-focused changes produced greater changes in memorability, and a 1-way ANOVA revealed significant differences in mean absolute changes in memorability between the rings (F(4, 1419) = 4.58, p = 0.0011)). Further, post hoc t-tests revealed that Rings 1 and 2 had significantly lower mean absolute changes than Rings 4 (p < 0.05; Q = 4.27, p < 0.05) and 5 (Q = 4.28, p < 0.05; Q = 4.27, p < 0.05). Since Rings 3, 4, and 5 contained the vast majority of center points, and resulted in much greater changes in memorability, center-focused changes, both on average and cumulatively, appear to have a greater impact on memorability than peripheral changes.

Figure 5: Average Change in Memorability Based on Center Point Location.

This figure shows the standardized 1080×1080 px image whereby the color of each ring represents the average change in memorability of center points located within the given ring. As shown in the figure, each concentric ring has a “width” of 108 pixels, with the total area of each ring decreasing towards the center. Most notably, Difference Images with center points located in the center ring (Ring 5) had, on average, had over twice the change in memorability than the outermost ring (Ring 1), suggesting that center-focused changes in an image change memorability more than peripheral changes.

4. DISCUSSION

Here, we revealed the spatiotemporal dynamics of how image memorability changes during the art creation process. From the start of the creation process, a piece’s fate in memory is largely set in stone–with the starting frame predictive of how memorable it will be at the composition stage and at its final completion. Pieces become more distinctive as they progress past their starting point. However, as time progresses, the impact of changes to a piece of art decreases. In particular, blocking and sketching produced the largest impacts on memorability, while the addition of details did not cause much of an impact. Interestingly, we found many visual features did not impact memorability; the visual perspective, color, spatial frequency, image integration measures, symmetry, or metabolic cost of a piece did not influence memorability. However, we did observe an impact of some visual features, where generally more heterogeneous but less cluttered images were more memorable. We also observed an impact of spatial decisions on memorability, where larger changes, particularly those central and lower in the image, resulted in the biggest changes in memorability.

Interestingly, the majority of our effects related to changes in memorability, but not a change in a specific direction. For example, making blocking and sketching changes in the piece, or changing elements lower in the image only impacted the absolute memorability (i.e., changed the memorability score of the piece), but did not consistently impact it in one direction (i.e., making it more memorable or forgettable). The only directional impact we observed over time was that more changed pixels from image to image resulted in a loss of memorability in the image. It is still an open question of why spatially bigger changes would cause a piece to become more forgettable, since the opposite end of the spectrum—smaller detailed changes—don’t appear to impact memory. We did also find some visual features that were correlated with memorability of the final piece; namely, higher anisotropy and sparseness, but lower self-similarity and clutter were correlated with higher memorability (and note that these measures were all correlated with each other, as well). These results echo findings from prior work suggesting that memorable images tend to be those that are simpler and less cluttered (Kramer et al., 2023; Davis & Bainbridge, 2023). Additionally, these changes resemble those made by the GANalyze generative neural network to improve the memorability of an image; it generally removes clutter and focuses in on the central object to make it more memorable (Goetschalckx et al., 2019c). However, as a whole, we still have not found a “holy grail” of prescriptive steps that will make a piece of art unforgettable.

These findings hold strong implications for artists, emphasizing the importance of the early stages in forging a memorable piece. Artists can thus test the memorability of their piece before finalizing it, to get a sense of its impact early on. In fact, if one wants to create a more memorable piece, they should stop sooner rather than later: changes later in the art creation process have less of an impact on the memorability of the piece. Additionally, while the cliché says that “The devil is in the details,” we demonstrate here that it is certainly not the case when it comes to an image’s memorability—it is the larger changes associated with blocking and sketching that artists should focus on.

We also discovered that the neural network ResMem was significantly able to predict the memorability of the different steps of the art creation process. It could successfully ascribe a memorability score to the composition and final stages of each drawing, correlating with true human memory performance. This predictive ability is impressive, given that ResMem’s training set consists mostly of photographs of the real-world (Needell & Bainbridge, 2022). It is thus not surprising that its predictions correlated best with human memory for the final piece, which is likely most similar to the full images in its training set. However, the fact that ResMem can also predict memory for the relatively sparse composition stage shows that there are unifying principles that govern people’s memories across a range of image types. Also, while ResMem’s predictions for the starting images did not correlate with human memory performance, its predictions for the other stages could predict memory performance for the starting image. This again highlights how an artwork’s memorability is set from the beginning, but also shows that ResMem may falter for the earlier planning stages of a piece. However, once a piece is composed, artists and creators can use ResMem as a validated tool to estimate the memorability of a piece as they create it. This presents interesting opportunities for artist-AI collaboration, but also inspires broader conversations about the goals of art, and the consequences of being able to intentionally create “unforgettable art.”

Our study also has important implications for our understanding of how images impact memory. Our findings reveal that it is really the bare bones—the composition, or the overall meaning of an image—that impacts one’s memory, not the small details. Indeed, we found that low-level visual features (e.g., color, spectral energy, perspective) played little role in what people remember, with a stronger influence of more compositional mid-level features, like the clutter or heterogeneity within a piece. These results dovetail nicely with other work with photographs showing that memorable images tend to be more easily categorizable when they are viewed as small thumbnails, compared to forgettable images (Goetschalckx et al., 2019a). In other words, it’s really the broader gist of the piece that determines its eventual memory fate, and gists that are easier to capture better stick in memory. We also found that certain locations in an image may be more spatially privileged in memory—especially central changes and changes lower in the image. We think this is likely due to how our images are spatially constructed—often consisting of a horizon, ground, and ceiling/sky, and with key information usually concentrated centrally. It is possible that when changes occur lower in the Y-axis (Figure 4), closer to the foreground and main subjects of many paintings, they are more noticeable, producing a greater impact on how the image is interpreted, and thus, having a greater impact on its memorability. There is ongoing computational work to develop neural networks that can make memorability heatmaps for images—indicating which regions are more or less memorable (e.g., Fajtl et al., 2018; Kyle-Davidson et al., 2025), so in the near future, neural networks could show what regions of a piece of art are particularly memorable or forgettable, helping artists to know where to edit their pieces.

We also tested the relationship of several aesthetics principles with the memorability of these pieces of artwork. Several measures that have shown correlations with aesthetic ratings, such as image integration (Nara & Kaiser, 2024), symmetry (Damiano et al., 2023), and metabolic cost (Tang et al., 2025) did not show a relationship to memorability. While we did not collect aesthetic ratings for the pieces here, prior studies have shown no correlation between the aesthetics of a photograph (Isola et al., 2013) or piece of artwork (Davis & Bainbridge, 2023) and its memorability. However, we did observe relationships to some other features related to balance, sparseness, and simplicity of the piece (Redies et al., 2023). Future work will thus need to examine how aesthetics relates to memory, and how other aesthetic features may impact the memorability of a piece. For example, spatial principles such as inward bias and visual balance could potentially impact the memorability of a piece both as a final version and as it is created (Palmer et al., 2013). The impact of Gestalt principles such as leveling and sharpening that occur during the detail stages of the creation process may also potentially impact the memorability of an image in different ways, especially given that memorable images may be a mix of both prototypical (i.e., level) and distinctive (i.e., sharp) elements (Goetschalckx et al., 2019a; Van Geert et al., 2025). Finally, while we looked at changes as a whole in this study, future work can separate changes as either additive or subtractive changes, to test their differential impacts.

Another key question that this work brings to mind is that of intentionality. Given that these artworks were collected from videos often serving as tutorials or just a display of the artist’s process, it is highly unlikely that the artists had the memorability of the piece in mind as they were creating it. However, a recent study tackled this very issue, where artists were explicitly challenged to create what they thought would be memorable and unforgettable artwork which were then shown and tested in an in-person gallery (Davis, et al., 2025). The researchers found that artists could intentionally create memorable or forgettable pieces of art, even though the average person could not predict what was memorable. However, surprisingly ResMem predicted people’s memory better than artists and laypeople could. This study demonstrated that memorability is a construct that can be deliberately guided through artistic intuition but is even better accounted for by computational models. This may suggest a path forward for artist-AI collaborative artwork, where artists can leverage AI tools to create the most memorable pieces.

Additionally, this work has strong implications for neuroaesthetics, which seeks to understand the influence of the arts and aesthetic experiences on the mind. Most notably, it directly relates to multiple foundational theories in neuroaesthetics like Leder’s model of aesthetic experiences (Leder, 2004), Tinio’s Mirror Model of Art, and the Aesthetic Triad (Chatterjee & Vartanian, 2014). Leder’s model seeks to detail the many steps that lead to one’s experience with an object of aesthetic interest, including perceptual analyses, classification, evaluation, context, and more. Most relevant, it asserts that implicit memory integration contributes to automatic continuous affective evaluation. Our work similarly suggests that the memorability of the piece as it is created may play a role in the decision making of the artist creating it. Tinio’s model interprets aesthetic experiences as a mirror to the art making process itself. A key component of the model is memory-based processing, that is meant to reflect the expansion and adaptation component of the art making process. This is especially relevant to this study given that our goal was to observe the impact of the art making process and memorability on one another. Finally, the Aesthetic Triad (Chatterjee & Vartanian, 2014) brings a much more liberal approach, stating that aesthetic experiences come about from interactions between three systems: Sensory-motor, Knowledge-meaning, and Emotion-valuation (Chatterjee & Vartanian, 2014). While it does not explicitly mention the role of memory as Leder’s model does, the Aesthetic Triad lists “context” as a subfield of Knowledge-meaning. Context plays a large role in memory, and has been most recently identified to play a key role in the reconstruction of long-term memory (Easton et al., 2024). Memory reconstruction has been directly linked to memorability, with reconstruction error being linked to stronger memory (Lin et al., 2024). Thus, overall, we propose that memorability may play an important role in both the art creation process and the aesthetic experience, and could potentially present an interesting update to these models.

We hope that the many insights that we reveal surrounding memorability are beneficial to both artists and researchers alike. However, we recognize that this investigation into an artwork’s creation process is just a start, and that there are many more avenues to explore. For example, future work could look at the semantic changes that occur with artistic decisions (e.g., how does the meaning change of a piece, or what are the specific objects that are inserted / deleted) to see if those are more predictive of changes in memorability. We also think the current methods serve as an exciting approach to use with other visual perceptual questions beyond just those about memory. Specifically, we have presented a “bottom-up approach” where we examine how the perception of an image is built, element-by-element, rather than the traditional “top-down approach” where we must tease apart the contribution of elements within a completed image. These methods could be used to answer questions about how abstraction occurs during visual perception, or how aesthetic impressions change frame-by-frame. This technique could also be used to investigate the creation process for other creative media beyond digital artwork, including paintings, sculpture, music or writing. Lastly, while this study focused on the goal of creating memorable pieces, this is not the sole goal that artists have when creating a piece. Future work could instead observe artists during the creation of pieces tasked with differing goals (e.g., to make a piece memorable, beautiful, emotional, or meaningful), to test how these goals modulate the dynamic art creation process. Research could also observe the impact of different artistic decisions given the same starting point or template.

Overall, the artist has the power to manipulate the memory of the viewer, through the decisions they make as they create a piece. These results suggest that certain choices will have a greater impact on that memory (e.g., how they compose the piece), and that technology such as artificial intelligence can measure this impact along the way.

HIGHLIGHTS.

• The memorability of artwork is largely determined from the beginning.

• A neural network can predict human memory throughout the art creation process.

• Overall simplicity rather than smaller artistic details influence art memorability.

• Blocking and sketching, particularly lower in the image, impact memorability.

Data Availability

Links to videos, ResMem- and human-derived memorability scores, and our artist-coded, visual, AI-based, and spatial metrics is available on the Open Science Framework (OSF) at https://osf.io/9z5rj/. A document with links to the available artwork creation YouTube videos is also available on our OSF repository.