stable-diffusion-xl-1024-v1-0 image-to-image/masking generating hard-edged results

[fxtech]

We're noticing with the released XL model that when passing in a soft edge matte, the generated inpaint/outpaint has a hard edge right along the 50% transparency point on the mask. The beta model didn't do this.
Is there something new in the API we need to be mindful of to avoid this? Seems to happen regardless of mask_source setting.