Repository metrics
- Stars
- (1,958 個のスター)
- PR merge metrics
- (PR metrics pending)
説明
This is a feature request to track a potential PR to implement it. I'm also looking for feedback for both naming and implementation preference.
Users can currently speed up Target Quality by only encoding down to only 1/4 of the frames in a chunk by specifying --probing-rate 2..4. We'll call this current behavior the "rate" strategy. We find this strategy acceptable because we consider that not all frames in a scene/chunk are necessary to accurately probe it and the speed-accuracy tradeoff is worth it. With that in mind, I propose a different strategy for reducing the computation in Target Quality probing: the "subset" strategy. This strategy would probe only the middle N frames in a chunk. Users can specify the number of frames to probe with --probe-strategy subset --probing-subset N. So for a given chunk of 400 frames, users can specify --probing-subset 50 to probe only 50 frames in the middle.
It's possible to implement both strategies simultaneously but for simplicity we can only support just one strategy at a time. For now, I initially propose 2 new options: --probe-strategy (or alternatively --probe-type) and --probing-subset <INT>. The value for --probe-stategy can be rate or subset and defaults to rate to match the current behavior. --probing-rate <INT> will be ignored when using --probe-strategy subset and vice-versa for --probing-subset <INT> and --probe-strategy rate (default).
I understand there are valid reasons to avoid using subsets in Target Quality mode, but I have had satisfactory results with this strategy in my own custom reimplementation of Target Quality and would like to see it implemented in the core product as an optional feature. I personally use ~20 frames per chunk so it's significantly faster than 1/4 of a very long chunk.
Thanks, - Boats M.