Introduction
In the field of vision science and cognitive psychology, researchers have long grappled with a fundamental trade-off: ecological validity versus experimental control. Ecological validity requires complex, real-world stimuli that reflect the visual richness of the environment, while experimental control demands the ability to precisely manipulate specific features. Traditionally, researchers used simple geometric shapes or color wheels to achieve control, but these fail to capture the complexity of natural vision. Conversely, using natural photographs often lacks a quantifiable metric for "distance" between stimuli.
This paper presents a novel pipeline leveraging Stable Diffusion XL (SDXL) to bridge this gap. By generating high-resolution, naturalistic image sets that vary along a smooth perceptual continuum, the authors provide a tool for studying vision and memory with stimuli that are both ecologically valid and parametrically controlled.
Methodology: The Synthesis Pipeline
The researchers developed a three-stage workflow to create and validate their stimulus library.
1. Latent Interpolation and Selection
Using SDXL, the team generated 60 variations of specific objects (e.g., a "fish") by maintaining a constant text prompt while varying the starting noise. To establish the endpoints of a perceptual scale, they utilized the Learned Perceptual Image Patch Similarity (LPIPS) metric—a deep-learning-based measure of visual similarity—to select two distinct images: an "anchor" and a "guide."
2. Spherical Linear Interpolation (Slerp)
To create a smooth transition between the anchor and the guide, the authors performed Spherical Linear Interpolation (Slerp) within the latent noise space. Unlike standard pixel blending, which often results in "ghosting" or semantically nonsensical intermediate states, interpolating in the latent space allows the model to generate coherent, high-resolution (1024x1024) images at every step. From 200 generated intermediate steps, 10 images were selected that were roughly equidistant according to the LPIPS model.
3. Psychophysical and Memory Validation
To ensure the computational "perceptual distance" aligned with human experience, the authors conducted a "method of triads" experiment involving 1,113 participants. Furthermore, they validated the utility of these images in a working memory task (delayed match-to-sample), testing whether the perceptual distance between two images predicted the likelihood of participant confusion.
The Stimulus Dataset
The resulting dataset, which has been made publicly available, consists of:
- 108 unique object scenes across six categories: Animals, Plants, Landscape Elements, Buildings, Items, and Vehicles.
- 10 variants per scene, ordered along a perceptual continuum.
- 1,080 high-resolution images in total.
Key Insights
Latent vs. Semantic Space
A critical finding of this work is that by keeping the text prompt fixed while interpolating in the noise latent space, the model modifies the perceptual features of an object without altering its semantic identity. This avoids the "uncanny valley" effects often seen in traditional morphing techniques, where one object morphs into a different category entirely.
AI as a Proxy for Human Vision
The study found a strong alignment between the LPIPS metric and human similarity judgments, with a Spearman’s correlation of $\rho = 0.73$. This suggests that the internal representations of deep neural networks are increasingly effective proxies for human psychophysical scaling, allowing researchers to automate the creation of perceptually grounded stimulus sets.
The Challenge of Perceptual Jumps
Despite the high resolution of the latent space, the researchers noted "perceptual jumps"—instances where a minor change in the latent noise resulted in a disproportionately large visual change. To maintain a smooth gradient, approximately 12.4% of the images required manual screening and replacement.
Results and Discussion
The psychophysical validation confirmed that human similarity judgments largely mirrored the AI-generated sequences. In the working memory tasks, participant accuracy followed a predictable graded pattern: accuracy was highest when the perceptual distance between the target and the distractor was large ("Easy" condition) and lowest when the distance was small ("Hard" condition). Interestingly, while participants showed learning effects over time, the underlying "confusability" gradient remained stable, indicating the robustness of the stimulus set.
Limitations and Future Directions
While the pipeline represents a significant advancement, several limitations remain. The model occasionally produces "hallucinations" or anatomical artifacts (e.g., animals with extra limbs), requiring manual quality control. Additionally, the current method does not allow for specific dimensional control (e.g., changing only the color or lighting); the variations are naturalistic but holistic. Future research may focus on scaling these similarity metrics across different object categories and developing methods for more granular feature manipulation.
Conclusion
By combining the generative power of Stable Diffusion with rigorous psychophysical validation, this research provides a powerful new framework for cognitive science. The resulting 1,080-image library offers a standardized, high-resolution resource for investigating the complexities of human perception and memory in a way that was previously unattainable with traditional stimuli.
Summary updated