Defining and Quantifying Creative Behavior in Popular Image Generators

TL;DR

The paper tackles defining and quantifying creativity in popular img2img generators by introducing a chain-based framework that measures three core facets: prompt-satisfaction, cohesion, and diversity, culminating in a Creativity Ranking $CR$. It formalizes these notions with mathematical definitions and applies them through iterative chains across multiple models and temperatures to reveal how image input strength and textual guidance shape creative outcomes. Empirical results show statistically significant effects of chain type and strength, with textual input proving essential for meaningful creativity and image-only inputs offering limited novelty in many cases. The work provides a practical, deployable set of metrics that help users select models and settings for task-specific creative needs and suggests avenues for improving model design and evaluation in the future.

Abstract

Creativity of generative AI models has been a subject of scientific debate in the last years, without a conclusive answer. In this paper, we study creativity from a practical perspective and introduce quantitative measures that help the user to choose a suitable AI model for a given task. We evaluated our measures on a number of popular image-to-image generation models, and the results of this suggest that our measures conform to human intuition.

Figures (5)

• Figure 1: Example Flux chains, generated from a single seed image apple_pie_139.png (the leftmost image), with different input temperatures.
• Figure 2: The flowchart depicting a single step in the chain process
• Figure 3: Example img_capKandinsky 2.2 chains for various strengths, generated from a single seed image apple_pie_184.png. From left to right: the seed image, followed by steps 1-$n$ of the generated chain where $n$ is the last image that satisfies the prompt requirements.
• Figure 4: Example images generated at step $1$ of their respective img_only chains, all seeded with the same image. Flux in the top row, Kandinsky 2.2 in the middle row, and Stable-Diffusion 3 in the bottom row; strength $0.3$ in the leftmost column, $0.6$ in the middle column, and $0.9$ in the rightmost column.
• Figure 5: Example images generated at step 1 of their respective img_cap chains, all seeded with the same image. Flux in the top row, Kandinsky 2.2 in the middle row, and Stable-Diffusion 3 in the bottom row; strength $0.3$ in the leftmost column, $0.6$ in the middle column, and $0.9$ in the rightmost column.

Theorems & Definitions (3)

• Definition 1: Satisfaction of Prompt Requirements
• Definition 2: Cohesion
• Definition 3: Diversity