Work

View-Conditional Representation Concentration in AI-Generated Anime Character Images: A Multi-Encoder Study of Civitai

Year: 2026

Type: article

Abstract: Assessments of visual homogenization in generative AI image communities may depend on the image view, visual encoder, and preprocessing procedure. This study develops a view-conditional representation concentration (VCRC) framework and applies it to a frozen Civitai collection comprising 1099 full images and 1096 face-cropped views. DINOv2 embeddings were evaluated using the Vendi ratio, mean pairwise cosine similarity, and average 10-nearest-neighbor similarity, with regional, resolution, and padding sensitivity analyses. Face-cropped views showed a lower Vendi ratio than full images (0.019120 vs. 0.024360) and higher mean pairwise similarity (0.490709 vs. 0.438244). The DINOv2 concentration effect remained positive in the 1088-image audited paired sample under the tested padding and resolution controls and in a 176-image subset whose crops met the primary-face coverage criterion under two independent annotations. In the full audited sample, OpenCLIP and ConvNeXt showed the same effect direction, whereas MAE showed the opposite direction; the magnitude of the MAE contrast was strongly affected by padding. In a separate 205-image annotation audit, the crop contained at least 90% of the primary face in 86.3–87.8% of images, depending on the annotator. Platform tags were not verified against generation prompts, and a matched hand-drawn baseline was unavailable. VCRC therefore provides a reproducible, encoder-aware framework for characterizing view-dependent representation concentration and testing its sensitivity to image processing. Future work should assess its transfer across platforms, generative models, and matched non-generated image collections.

Source: Applied Sciences

DOI: 10.3390/app162010279