I've posted a new video to YouTube showing some of the speed improvements I've made recently.
Since the last video in Feburary I've moved the entire voxel -> mesh pipeline (except for the seam mesh generation) to the GPU via OpenCL and reworked how the volume is processed. I'm planning on a blog post to explain the changes but the main thing is that I no longer always sample the field at a 1:1 resolution. E.g. for a LOD1 node I now only sample a 128x128x128 volume instead of 8, and for LOD2 only 1 volume instead of 64. This has had a huge impact on the performance which has helped with other features, like the dragging operations you can see in the video.