It only has a Vulkan implementation actually! So if it works at all then it is certainly using Vulkan. The memory costs and performance is very dependent on the resolution of the input frame, the more powerful the GPU the better the performance. Otherwise, it will be pretty slow... It will have to do a lot of redundant non-sharable work each frame at the moment until I can dial in on some optimizations unfortunately. For example, with still images there is a chance that work can be shared and re-used between frames in parallel. Though, for video-footage where each frame is basically a completely unique image, it will have to uniquely seam-carve each and every frame, which is pretty hard to avoid at the moment.