We originally estimated 70% accuracy, but survey-based fit predictions outperformed, averaging 77.4%—a win beyond our initial goal.
Despite aiming for 95% accuracy, body scan predictions averaged 85.7%, leaving room for improvement.
Midway through testing, we found that avatars were being generated with incorrect shoulder widths, directly skewing sleeve predictions.
Since garments have a fixed waist position, individual torso length affected how the waist actually fit. This finding led the team to consider integrating torso height into future predictions.
Some participants preferred a snug, form-fitting look, while others leaned towards a looser, more relaxed fit—making it really hard to get a success rate - proving that fit is more personal than just numbers.
👱🏻♀️ "I’d buy the 2X because even though it’s a little, baggy, it’s more comfortable this way. I like my shirts to be on the larger side"
👩🏽🦱 "I prefer my clothing to fit tighter, so I usually choose smaller sizes to avoid excess fabric.”
Users found terms like "tight" and "loose" confusing since they’re not commonly used in fashion.
While we didn’t have exact metrics, confusion noticeably dropped.
Users describe the avatars as bland, clinical, and lacking personality, comparing it to "naked mannequins" or a "TSA scan."
👩🏼🦰 “Nothing excites me about this avatar. It reminds me of a medical app.”
To explore a better alternative, I used AI and Photoshop to create avatars that feel more engaging and approachable—without being uncomfortably realistic.
These explorations aim to breathe life into the avatars because shopping should be fun, not bland and clinical!
