In natural vision, high-level features such as global scenes and discrete objects consistently covary with lower-level statistical constituents including textures, local surface properties, and edge contours. For cognitive neuroscientists and vision researchers, this systematic real-world entanglement creates a persistent interpretive hurdle: when observing human behavior or neural responses to complex natural scenes, it is often challenging to determine whether an effect reflects sensitivity to semantic categories, such as a jungle or kitchen, or to the specific low-level edge distributions and surface textures that naturally define those environments. To address this fundamental experimental compromise between well-controlled but artificial scrambled stimuli and ecologically rich but entangled natural photographs, a research team developed the SPOT framework. By systematically crossing settings, patterns, objects, and textures within a generative diffusion pipeline, the researchers produced naturalistic synthetic images that allow experimental control across distinct tiers of the visual hierarchy.
The methodological foundation of the SPOT pipeline relies on the latent diffusion architecture Stable Diffusion XL, specifically utilizing the pre-trained SDXL 1.0 base model alongside its default refinement model for final denoising. Standard text-to-image synthesis models are inherently designed to generate coherent semantic scenes from textual prompts but lack granular mechanisms for independently manipulating low-level structural and stylistic properties. To achieve precise experimental intervention, the researchers augmented the generation pipeline within the Automatic1111 web user interface using two modular extensions: Low-Rank Adaptation, or LoRA, and ControlNet. While high-level settings and objects were specified directly through weighted prompt components and occasional negative prompts, low-level surface textures were controlled by fine-tuning SDXL cross-attention layers via specialized LoRA models. Simultaneously, low-level spatial edge distributions were imposed by using ControlNet conditioned on Canny edge detection maps derived from standardized geometric templates.
To establish a tractable stimulus space while proving the viability of factorial visual manipulation, the investigators structured their stimulus library around two paired feature axes. High-level feature pairs combined a specific setting with a dedicated primary object, such as pairing an underwater setting exclusively with a teacup or an open field with a truck. Low-level feature pairs combined a specific surface texture with an edge pattern, such as pairing a paper texture with diamond contours or crayon strokes with horizontal stripes. The team initially generated an extensive library of 1,594 images spanning a 10-by-10 grid comprising 100 distinct four-feature groupings, rendering every image at a standardized resolution of 1024 by 1024 pixels using deterministic sampling parameters like Euler Ancestral or DPM++ 2M Karras. Hyperparameters, including classifier-free guidance scales and extension weights, were systematically tuned so that pattern outlines and stylistic textures were expressed strongly without degrading semantic object identity.
Following initial human behavioral pilot evaluations, the research team refined the full stimulus collection into an optimized 8-by-8 grid by dropping feature combinations associated with lower identification accuracy. From this refined pool, an intermediate set of 484 images was retained by removing exemplars that yielded misclassifications or prolonged reaction times exceeding two standard deviations above the grand mean. For rigorous empirical and computational validation, the authors selected a representative subset of 64 images, containing exactly one randomly chosen exemplar from each cell of the 8-by-8 grid. Because all generation metadata—including random seeds, sampler configurations, guidance values, and conditioning maps—remain embedded within each image file, every stimulus within the resulting SPOT Grid is reproducible and capable of controlled parametric modification in identical user interface environments.
To objectively confirm that the image synthesis pipeline modulated visual representations with appropriate hierarchical specificity, the authors tested the 64-image subset across four complementary computational models. Low-level visual properties were modeled using the EarlyVis neural network from the img2fmri package, which predicts functional magnetic resonance imaging BOLD responses in human early visual areas V1 and V2, alongside an analytical Gabor wavelet pyramid capturing multi-scale orientation and spatial frequency energy. High-level semantic features were evaluated using two distinct multimodal vision-language architectures: the contrastive dual-encoder CLIP paired with Sentence-BERT sentence vectorization, and the generative autoregressive vision-language model Florence-2 running caption generation before Sentence-BERT encoding. Pairwise cosine similarity matrices were computed across image pairs that shared low-level features only, shared high-level features only, or shared neither category of features.
Permutation-based significance testing over 10,000 iterations revealed a robust double dissociation across the computational models, confirming that targeted feature levels were successfully decoupled. In the EarlyVis model, representational similarity was significantly higher for image pairs sharing low-level textures and patterns than for pairs sharing high-level settings and objects, exhibiting an observed mean difference of 0.151 with a Bonferroni-corrected p-value below 0.0001. The Gabor wavelet pyramid demonstrated an identical low-level preference, with similarity significantly greater for shared low-level pairs than shared high-level pairs by a margin of 0.109. Conversely, high-level models exhibited the inverse representational profile: CLIP representations yielded substantially higher similarity for shared high-level features than shared low-level features, with a difference of 0.180, while Florence-2 demonstrated an even larger high-level advantage of 0.271, with all comparisons reaching significance at p below 0.0001.
Following computational validation, the researchers evaluated the perceptual recognizability of the manipulated features in a controlled behavioral experiment involving 30 human participants recruited via Prolific. Because naturalistic settings and common objects possess familiar semantic labels whereas fine-grained artistic styles and geometric edge templates do not, participants completed a brief label-training phase before testing. This procedure included viewing standardized training exemplars depicting an unstudied rooftop setting and cat object rendered in each targeted texture and pattern, supported by magnified instructional videos and animated edge overlays to ensure classification reflected genuine perceptual discrimination rather than vocabulary ambiguity. Participants were required to pass a timed pre-test criterion before advancing to the primary classification phase.
In the main experimental task, participants evaluated all 64 images across four blocked judgments, categorizing the setting, object, texture, and pattern of each stimulus under a 10-second response window without feedback. Overall labeling accuracy was high, yielding a raw grand mean of 96.8 percent across participants with a mean response time of 2.28 seconds on correct trials. Generalized linear mixed-effects modeling revealed that accuracy varied systematically across feature classes: object categorization achieved a model-estimated accuracy of 98.6 percent and pattern judgments reached 98.0 percent, whereas setting and texture judgments yielded model-estimated accuracies of 96.3 percent and 94.2 percent, respectively. Linear mixed-effects modeling on log-transformed latencies showed a geometric grand mean response time of 2.05 seconds, with object and pattern classifications executed significantly faster than judgments of background settings and surface textures.
Analysis of participant error patterns through confusion matrices demonstrated that categorical misidentifications were rare and largely unsystematic, with correct classification rates for individual features predominantly ranging between 95 and 99 percent. Cross-category confusions remained below 3 percent for nearly all feature pairs, with only a single notable exception: images rendered in the graffiti texture were labeled as oil paint in 8 percent of presentations. Further inspection indicated this elevated rate was primarily driven by three individuals who consistently confused the two styles; when those three participants were excluded in post-hoc analysis, the confusion rate fell to 3 percent. These behavioral metrics confirm that the synthetic stimuli remain perceptually intelligible and that human observers can rapidly and independently parse high-level semantic identities and low-level visual statistics within the generated scenes.
The authors emphasized that while the SPOT framework offers an adaptable tool for investigating hierarchical visual processing, feature binding, and cognitive representations, complete physical and perceptual decoupling between high- and low-level visual features remains inherently impossible. Natural semantic categories are partially defined by their low-level structural components, meaning a jungle cannot be depicted without some degree of foliage-like edges and organic textures. Furthermore, intermediate visual properties, such as three-dimensional volumetric contours, coarse spatial geometry, and functional object affordances like containment or graspability, inevitably remain tied to recognizable objects. Nevertheless, by providing a validated, publicly accessible stimulus library and an open generative framework, the authors demonstrate that modern diffusion models can bridge the traditional gap between experimental control and ecological realism across visual neuroscience and psychological research.
Subject of Research: Systematic manipulation and partial decoupling of hierarchical visual features in naturalistic synthetic imagery
Article Title: The SPOT Grid: Naturalistic images with predefined settings, patterns, objects, and textures
Article References: Hirschstein, Z., Aly, M., & Baldassano, C. (2026). The SPOT Grid: Naturalistic images with predefined settings, patterns, objects, and textures. Behavior Research Methods, 58(12), Article 336. https://doi.org/10.3758/s13428-026-03211-w
Image Credits: AI Generated
DOI: 10.3758/s13428-026-03211-w
Keywords: visual perception, feature hierarchy, Stable Diffusion, ControlNet, LoRA, computational modeling, visual cognition, image categorization, early visual cortex, scene perception, cognitive neuroscience, computer vision, image synthesis, perceptual scaling, behavior research methods, feature binding
Cite Scienmag News
Glenn Wilkins. (November 21, 2026). Generative Framework SPOT Partially Decouples Visual Features in Naturalistic Images. Scienmag. https://scienmag.com/generative-framework-spot-partially-decouples-visual-features-in-naturalistic-images/
Glenn Wilkins. "Generative Framework SPOT Partially Decouples Visual Features in Naturalistic Images." Scienmag, 21 November 2026, https://scienmag.com/generative-framework-spot-partially-decouples-visual-features-in-naturalistic-images/. Accessed 21 November 2026.
Glenn Wilkins. "Generative Framework SPOT Partially Decouples Visual Features in Naturalistic Images." Scienmag. November 21, 2026. https://scienmag.com/generative-framework-spot-partially-decouples-visual-features-in-naturalistic-images/








