Perceptual Evaluation of Garment Drape Intensity in AI-Generated Garment Images

Recent advances in generative artificial intelligence have enabled fine-grained manipulation of visual attributes in garment images, but consistency in automated feature space does not necessarily guarantee consistency in human perception. This study investigates the perceptual validity of garment drape intensity control in AI-generated garment images using a semi-automated Human-in-the-Loop evaluation pipeline. Five base garments were manipulated across five latent-space control levels and evaluated independently using Contrastive Language–Image Pre-training (CLIP)-based feature-space analysis and a two-alternative forced-choice (2AFC) experiment with 15 participants. The CLIP projection scores increased monotonically with the manipulation parameter for all five garment sequences. Human judgments showed a similar monotonic tendency at the aggregate level; however, garment-specific Bradley–Terry analysis revealed local reversals in the estimated perceptual ordering for three of the five base garments. These results demonstrate that aggregate agreement between automated and human evaluation can mask stimulus-specific perceptual discrepancies. The proposed framework provides a practical approach for combining automated screening with human perceptual validation when assessing fine-grained attributes in generative fashion systems.

@article{Kuwahara2026,
  title     = {Perceptual Evaluation of Garment Drape Intensity in AI-Generated Garment Images},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {9},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Noriaki Kuwahara and Yuta Yamamoto and Dongeun Choi and Kazunari Hirakoso and Takashi Satou},
  doi       = {10.14569/IJACSA.2026.0170940},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170940}
}