There is a question that gets asked a lot about children’s books: are there enough characters who look like me? It is the right question to start with, but it is not the only one. A book can feature a child of color as the protagonist and still, through dozens of small compositional decisions, render that character as less important than the story says they are. Smaller on the page, pushed to the upper edge and farther from the reader’s eye, drawn with less visual detail than the White characters surrounding them.

In 1990, Rudine Sims Bishop introduced the foundational metaphor that still anchors this conversation: books as mirrors, windows, and sliding glass doors. What the metaphor makes plain, though rarely made explicit in subsequent scholarship, is that the mirror is fundamentally a visual argument. When a child of color looks into a picturebook mirror, they do not just see a character who shares their identity — they also see how that character is visually presented: whether central or peripheral, large or small, richly detailed or simply drawn.

Before getting into what I found, a word on scope. The sample is small — 45 books — and it is skewed by design. Because I drew from award pools that have shifted heavily toward CoC protagonist books in recent years, 80% of the corpus features characters of color as protagonists. That means the White character sample is simply smaller — 93 observations compared to 245 for CoC — which affects the statistical precision of every comparison. Bounding-box annotation was performed by a single coder without intercoder reliability testing, and the visual metrics, while principled and validated, are proxies. Edge density captures structural rendering complexity, not artistic quality or emotional resonance. Foreground proximity rests on a simplified model of pictorial space that does not account for every compositional convention. The findings characterize this specific corpus. They are not a verdict on the industry, and I do not treat them as one.

I built a corpus of 45 picturebooks recognized for illustration excellence between 2015 and 2024, drawn from the Caldecott Medal and Honor, the Boston Globe–Horn Book Award for Picture Books, and the New York Times/NYPL Best Illustrated list. Three spreads were sampled per book — an early, middle, and late spread — to capture compositional diversity. I manually drew bounding boxes around 346 individual character observations across 141 spreads, tagged each character with their race, narrative role, and other coding fields, and then ran a computer vision pipeline to extract five visual measurements from every bounding box. By specific racial group, the CoC sample comprises Black (n = 132), Latino/Latina (n = 44), Asian (n = 35), Middle Eastern (n = 21), and Indigenous (n = 19) characters.

The Five Things I Measured

What I Found: Not a Simple Story

The honest answer is that the results are more complicated and more interesting than I expected, with different metrics telling different parts of the story. The data do not point to a single, consistent pattern of advantage or disadvantage — they point to a more differentiated structure in which compositional prominence and rendering detail can diverge.

The Numbers, in Brief

Not All Characters of Color Are the Same

One of the things I found most interesting, and most absent from the existing literature, is how much variation there is within the CoC category. A binary CoC/White comparison conceals substantially different patterns across groups.

Asian characters are the most centrally positioned of any group (centrality M = .256, the lowest and most central value) and have the highest visual complexity scores among CoC subgroups (edge density M = .149), approaching the White character mean of .137. This pattern likely reflects the specific illustrative traditions in the Asian-protagonist books in this corpus — particularly Watercress (Jason Chin, highly detailed watercolor), A Different Pond (Thi Bui, intricate panel compositions), and A Big Mooncake for Little Star (Grace Lin, decorative folk art) — rather than any systematic editorial preference.

Indigenous characters present the starkest compositional profile: the smallest mean bounding boxes (bbox M = .043, versus .096 for Black characters), the highest size rank (M = 2.33, meaning they are on average the third-largest figure on their spreads), and the highest edge density (M = .158). This combination of small compositional footprint with high rendering complexity is distinctive and concentrated in two books by Michaela Goade (We Are Water Protectors, Berry Song), where the human figure is embedded in sweeping environmental compositions rather than dominant over the visual field. This compositional choice may reflect deliberate relational aesthetics — a rejection of the Western picturebook convention that foregrounds the individual protagonist — rather than marginalization.

Latino and Latina characters score lowest on visual complexity (M = .092), a finding that warrants further examination in a larger corpus. Black characters, as the largest subgroup (n = 126), show the widest variance across metrics, reflecting the breadth of illustrative traditions represented: from Ekua Holmes’s collage-based work in Voice of Freedom to Bryan Collier’s watercolor in Trombone Shorty to Kadir Nelson’s oil paintings in The Undefeated.

A Decade in View: Trends Over Time

The longitudinal picture adds texture — and caution. Year-by-year analysis of the CoC–White edge density gap shows notable variability across the decade. The years 2015 and 2017 show CoC characters actually scoring higher on visual complexity than White characters, associated with high-complexity Caldecott-recognized books in those years. From 2018 through 2020, White characters consistently received higher edge density scores. From 2021 onward, the gap narrowed and reversed in several years, suggesting the possibility of gradual improvement in rendering parity (slope = +0.004/year).

The important caveat: with typically 5–30 observations per year-group, none of these year-level trends reach statistical significance on Mann-Kendall testing (all p > .10). The longitudinal data are appropriately treated as descriptive — illustrating the pattern of the decade rather than confirming its direction.

What This Means for How We Talk About Representation

The field of children’s publishing has made real and measurable progress on presence. In this corpus, 80% of books feature CoC protagonists — a figure that would have looked very different ten years earlier, and one consistent with Koss (2015) and Martinez et al. (2016) documenting earlier gaps that have since narrowed. The distinction Hanlin (2021) draws between shallow diversity (numerical inclusion) and deep diversity (inclusion with prestige and narrative authority) is exactly the distinction this data explores computationally.

But the rendering gap tells a more complicated story. Even in a corpus selected for illustration excellence — books that committees of experts judged as the best-illustrated books of their year — White characters still receive more visually detailed rendering than CoC characters. And at the protagonist level, CoC characters are positioned farther from the reader’s embodied viewpoint, in the pictorial register that Kress and van Leeuwen (2006) and Nodelman (1988) identify as conveying distance rather than intimacy. Increased presence and centrality do not uniformly correspond to equivalence in rendering or pictorial proximity.

Rudine Sims Bishop wrote that a mirror does not merely show who is present. It shows how one appears. My data suggest that the how is lagging behind the who, and that measuring it requires tools that go beyond counting heads on pages.