Measuring Aesthetic Homogenization in Text-to-Image AI: Diversity Collapse in DALL-E 3 and Imagen 4

Text-to-image models produce strikingly similar outputs, but the scale of this homogenization has not been systematically measured. We introduce a reproducible methodology for quantifying aesthetic and cultural diversity collapse, operationalizing three distributional measures - effective support radius, diversity index, and effective dimension - in CLIP and DINOv2 embedding space. Applying it to 1,177 images from two leading commercial models, DALL-E 3 and Imagen 4, we find that within-prompt outputs collapse to about 10 of 768 effective dimensions, discarding over 98% of the available dimensions; that minority cultural representations are 17% less diverse than majority ones under CLIP (attenuating to about 8% under DINOv2; p < 0.001, Hedges' g = 1.51); and that the two independently developed models converge to cosine similarity 0.86 on identical prompts, well above the 0.64 mismatched-prompt baseline. A controlled ablation on open-source Stable Diffusion XL isolates the mean-squared-error objective from preference training: at the standard guidance scale, 91% of the majority/minority diversity gap is already present in a base model with no preference training, while classifier-free guidance monotonically compresses diversity. Homogenization is therefore primarily structural - a property of the training objective more than of human-feedback tuning. We release all code and embeddings.