- Blender 3.5.1
- NVIDIA 3080 RTX
- i7-12700F
- The video sequencer cache limit is set to 16384 (I have 32G of memory)
I looked at this post which gave me the suggestion of increasing the cache limit.
I created a scene to chromakey (green screen) this clip: https://pixabay.com/videos/subscribe-button-subscribe-73041/
I converted the clip to 30FPS with ffmpeg to match the VSE project. The clip is Codec: H264 - MPEG-4 AVC (part 10) (avc1) (video) and Codec: MPEG AAC Audio (mp4a) (audio) (matches the original).
The composite preview area is set to a very small part of the 1920x1080 clip. You can also see that edit performance is set to low, "OpenCL" is checked, and "Buffer Groups" is checked.
Here is the VSE. Proxy render is set to 25%. The preview area is set to this small space around where the composite scene is inserted.
When playing the preview, as soon as Blender gets to the inserted composite scene, the frame rate drops to 8 from the target of 30. It doesn't matter if the preview has played once or multiple times. It doesn't ever appear to cache in a way that improves the frame rate, which contradicts the post referenced above regarding caching.
For a large or long keyed project, this seems like it can't be the way it normally works/normal performance.
Of interesting note, when looking at the composite preview area in the compositing tab and playing that scene, I never see the composite viewer move. It only moves when I click on an individual new frame in the composite scene, at which point I see an indication of "compositing" in the info area at the bottom of the Blender window.
Other than pre-rendering the scene to an image sequence and inserting that, what can I do to improve the frame rate on the preview here? I'm assuming I have something wrong either with my composite/keying settings or something else.
Update: Pre-rendering the scene to an image sequence (with transparency) and inserting it resulted in marginally better performance, but it was still quite slow (~10fps) and would sometimes cache and sometimes not. I think the slowness may be because the output is a full 1920x1080 image that's mostly transparent.
