Slightly more dataset (200 images),

better prompt adherence in my testing