I've recently been grading some shows cut in Premiere Pro that contain a lot of retimed footage. My workflow is fairly standard - I've been exporting an XML from Premiere along with a flattened offline reference movie to check for resizing and frame offset irregularities. I've noticed that on retimed clips, the results of Premiere Pro's optical flow algorithm are usually superior to what I can achieve in Resolve. This really surprised me, as I have always assumed that Resolve would have superior results to traditional NLEs when it comes to any type of image processing. I've experimented with the various options in Resolve for quality and motion range, and the same shots retimed in Resolve have consistently displayed more visible artifacts than the rendered reference movie coming form Premiere Pro (rendered on the same system). The retimed source footage is 1080p60, but the speed changes are often not an easily divisible fractional speed like 50% or 25%. The projects are 30p, and the 60p source footage is often retimed to arbitrary speeds like 60% or 70%.
My workaround has been to export a flattened Quicktime movie from Premiere Pro with no filters and optical flow retiming baked in, then blade it up in Resolve and cut in the rendered Premiere Pro footage for any shots that have been retimed. It works fine, but obviously it would be faster to simply let Resolve do the retiming in situ.
Has anyone else had a similar experience? Are optical flow calculations in Resolve GPU or CPU based? Would different GPUs produce different optical flow results? The system I'm grading on is a 12-core 2.6GHz Mac Pro 5,1 with an ATI Radeon HD 5770.