Severe facial hallucinations on inverted poses (Orientation Bias)
Hi team,
I’ve been stress-testing the model on Non-Canonical Poses (gymnastics, yoga, inverted postures) and found a significant failure mode regarding orientation logic.
Specifically, when the subject is upside down (e.g., handstand/backbend), the model suffers from "Orientation Bias." It attempts to render facial features (eyes/mouth) with an "upright" alignment onto an inverted skull, resulting in severe anatomical hallucinations
It seems the model struggles to decouple the "face" tokens from the "gravity" logic.
My team at Repalto specializes in fixing these rotational edge cases. We can construct a "Rotational Robustness" benchmark (poses rotated at 90°, 180°, and 270°) to help you map where the facial generation breaks down.
Happy to send that batch over if you want to use it to fine-tune the next version. Just let me know.
Best,
Wahaj Barlas
Repalto Data Ops