zyddnys

Use PDCNet+ for occlusion mask

1

There's a similar project https://github.com/zyddnys/sd_animation_optical_flow that uses PDCNet+ for optical flow confidence, have you tried using PDCNet+?

Fix incorrect image channels when dealing with 1x1 image #8954

2

Since input is PIL image, the input data format is always channels last, this PR prevents transformers from inferring image format and thus fix #8954