TraitSpaces: Towards Interpretable Visual Creativity for Human–AI Co-Creation

Prerna Luthra · 2025

Read TraitSpaces: Towards Interpretable Visual Creativity for Human–AI Co-Creation by Prerna Luthra (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-24326.txt.

Introduction

Creativity resists numbers, yet shared terms could help artists and AI systems work together. Prerna Luthra introduces TraitSpaces, a framework for visual creativity grounded in psychology and shaped by interviews with practicing artists. It models creativity across four worlds: inner, outer, imaginative, and moral. Twelve traits capture affective, symbolic, cultural, and ethical facets.

Using twenty thousand artworks, Luthra labels images with a large model under theory-aligned prompts, then tests which traits can be learned from image embeddings. Some traits, like redemptive arc, predicted reliably; others, like memory imprint, resisted purely visual coding. A trait-space display lets users slide along a trait to explore works, supporting co-creation.

Catalog record

CREATORPrerna Luthra
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE18 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-24326

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

TraitSpaces: Towards Interpretable Visual Creativity for Human–AI Co-Creation Prerna Luthra Abstract We introduce a psychologically grounded and artist-informed framework for modeling visual creativity across four domains—Inner, Outer, Imaginative, and Moral Worlds. Drawing on interviews with practicing artists and theories from psychology, we define 12 traits that capture affective, symbolic, cultural, and ethical dimensions of creativity. Using 20k artworks from the SemArt dataset, we annotate images with GPT-4.1 using detailed, theory-aligned prompts, and evaluate the learnability of these traits from CLIP image embeddings. Traits such as Environmental Dialogicity and Redemptive Arc are predicted with high reliability ( R 2 ≈ 0.64 − 0.68 R^{2}\approx 0.64-0.68 ), while others like Memory Imprint remain challenging, highlighting the limits of purely visual encoding. Beyond technical metrics, we visualize a “creativity trait-space” and illustrate how it can support interpretable, trait-aware co-creation - e.g., sliding along a Redemptive Arc axis to explore works of adversity and renewal. By linking cultural-aesthetic insights with computational modeling, our work aims not to reduce creativity to numbers, but to offer shared language and interpretable tools for artists, researchers, and AI systems to collaborate meaningfully. 1 Introduction Creativity assessment in AI has largely relied on narrow proxies—fluency/originality scores from psychometric traditions [Torrance, 1974; Kaufman & Beghetto, 2009] or technical image metrics like FID [Heusel et al., 2017] and CLIPScore [Hessel et al., 2021]—that overlook affective, symbolic, cultural, and ethical dimensions. While broader theories frame creativity as emerging from person–process–product–context interactions [Amabile, 1983; Csikszentmihályi, 1996; Boden, 2004], operationalizing such frameworks for vision-language models remains challenging due to tensions between psychological richness and computational tractability. To bridge this gap, we synthesize psychological theory with artist interviews into a four-world framework (Inner, Outer, Imaginative, Moral). Its 12 traits (e.g., Emotional Intensity, Environmental Dialogicity) balance conceptual depth with partial computability, enabling application across creative domains. We focus on visual art—where these dimensions manifest concretely in composition using 20k SemArt images [García & Vogiatzis, 2018]. Using GPT-4.1 as a scalable annotation proxy [Tan et al., 2024], we score artworks per trait, then regress against CLIP embeddings [Radford et al., 2021] to: (1) Quantify visual encodability of creative dimensions, (2) Derive interpretable trait directions for semantic control (e.g., "increase Redemptive Arc"), (3) Propose Human AI co-creation tool framework based on trait directions. This establishes the foundation for trait-aware human–AI collaboration beyond novelty-centric evaluation. While our framework introduces computable dimensions, it is not intended to reduce creativity to numbers. Instead, it offers a shared language to connect artistic, cultural, and psychological insights with interpretable computational tools. 2 4-World Creativity Framework We developed this framework through semi-structured interviews with practicing artists and writers about their creative process, and then synthesized those insights with foundational psychological and creativity theories. Participants often described creativity as a form of exploration, of inner experience, the external world, imaginative possibility, and moral concern, an inductive theme that, together with prior theory, motivated our four ’world’ framework. The resulting taxonomy organizes twelve traits into four domains—the Inner , Outer , Imaginative , and Moral worlds—that reflect complementary facets of creative practice. Inner World. Human creativity also involves what we term the Inner World —a cluster of introspective dimensions grounded in affect, memory, and symbolic cognition. We operationalize this as three core components: Emotional Intensity , Memory Imprint , and Personal Symbolism . Emotional Intensity captures the raw immediacy and introspective depth of affective experience. Rooted in psychodynamic theory [Freud, 1908], intense emotion is sublimated into symbolic form, while narrative psychology emphasizes emotion as central to life-story construction [McAdams, 2001]. Memory Imprint reflects the autobiographical embedding of lived experience—drawing on sensory detail, flashbacks, and self-referential cues that ground the work in personal history [Conway & Pleydell-Pearce, 2000; McAdams, 2001]. Finally, Personal Symbolism involves a creator’s idiosyncratic system of metaphors, archetypes, and surreal logic, anchored in depth psychology [Jung, 1964]. These dimensions align with foundational theories of internal ideation in the Geneplore model [Finke et al., 1992], pointing to psychologically rich and authentic aspects of creative practice. Outer World. Beyond internal ideation, creative work is deeply embedded in the Outer World —shaped by culture, environment, and social discourse. We define this domain through three interrelated constructs: Cultural Situatedness , Environmental Dialogicity , and Social Reflexivity . Cultural Situatedness refers to how grounded a work is in a specific cultural, historical, or geographic context. Drawing from sociocultural theory [Vygotsky, 1978; Glăveanu, 2014], it reflects how creators engage with shared cultural tools, traditions, and domains. Environmental Dialogicity captures how the physical or natural environment participates in meaning-making, aligning with ecocritical frameworks that treat landscapes as active co-agents in narrative [Buell, 2005; Bate, 2000]. Finally, Social Reflexivity involves critical engagement with audience and power structures, grounded in dialogic and performative theory [Bakhtin, 1981; Butler, 1997], whereby creative works both reflect and challenge the norms of their social milieu. Together, these dimensions frame creativity as interaction across systems of meaning. Imaginative World. Creativity within the Imaginative World reflects a symbolic, surreal, and playful cognitive space characterized by three interrelated dimensions: Surreal Divergence , Symbolic Density , and Playful Subversion . Surreal Divergence captures the blending of reality with dreamlike or fantastical elements that subvert ordinary logic and perception. Rooted in psychoanalytic theory [Freud, 1919; Jung, 1964], this dimension evokes visionary, subconscious-driven experiences through otherworldly imagery that defies everyday expectations. Symbolic Density denotes multi-layered, richly woven symbolism, where motifs or events carry multiple, often ambiguous meanings, inviting diverse interpretations and deep engagement [Fauconnier & Turner, 2002; Eco, 1976]. Finally, Playful Subversion embodies the freedom to bend or break conventions—mixing genres and employing irony, absurdity, and novelty to surprise and engage the audience—resonating with postmodern aesthetics of disruption and hybridity [Hutcheon, 1989]. Collectively, these dimensions define a richly imaginative space where symbolic complexity and playful originality converge. Moral World. Creativity in the Moral World engages ethical, communal, and transformative dimensions, situating works as arenas for moral reflection and social change. We conceptualize this domain through Ethical Provocation , Collective Resonance , and Redemptive Arc . Ethical Provocation reflects a work’s capacity to confront audiences with complex moral dilemmas or injustices, provoking reflection without simple resolutions, consistent with Nussbaum’s notion of moral imagination [Nussbaum, 1990]. Collective Resonance denotes the articulation and amplification of communal experiences—particularly those of marginalized groups—fostering empathy and solidarity [Polletta, 2006; hooks, 1994]. Finally, the Redemptive Arc captures trajectories where adversity catalyzes healing, transformation, or hope, reflecting McAdams’s redemptive self-stories and Campbell’s universal hero’s journey [McAdams, 2006; Campbell, 1949]. Together, these facets frame creativity as ethical engagement, communal meaning-making, and social renewal. 3 Method Dataset: We use the SemArt dataset, a multimodal corpus of Western fine-art paintings (approximately 21,000 images) [García & Vogiatzis, 2018]. The training, test, and validation sets contain about 19,000, 1,000, and 1,000 images, respectively, as provided by the dataset. GPT-4.1 trait scoring: We operationalize the 12-trait rubric by passing detailed prompts. These prompts provide full trait definitions and a 0-4 integer rubric (Appendix B). For each image–trait pair, we query GPT-4.1 ( gpt-4-turbo-2024-04-09 ) with temperature =0 and top_p =1 , instructing it to return only an integer in \{0,1,2,3,4\} CLIP embeddings: Each image is encoded with CLIP ViT-B/32; we use the 512-D image embedding vector [Radford et al., 2021]. Ridge probe (interpretable trait axes): Let \mathbf{X}_{\mathrm{train}}\in\mathbb{R}^{N\times

Read the complete text