A tiling mosaic-esque video generator for the impatient with powerful GPUs.

You must already have CUDA, FFmpeg, cmake and standard gnu build-tools installed.

pip install pycuda numpy tqdm av
git clone https://github.com/NVIDIA/VideoProcessingFramework.git
cd VideoProcessingFramework
pip install .

Look at the help:

./mosaic.py -h

Example:

./mosaic.py -o myvideo.mp4 some/folder/with/videos/

If you want insane quality:

./mosaic.py -o myvideo.mp4 -ow 3840 -oh 2160 some/folder/with/videos/ -ea preset=P6 bitrate=70M profile=high

The code is already pretty fast. I tested with 16 20s 1440p 60fps clips, which took 40s to tile, so about 0.5x realtime. That amounts to a normalised speed of 8x realtime per clip, or a decoding performance of at least 480 frames per second (due to the reencoding overhead).

Although I would like even more speed, I doubt this is possible without rewriting the whole thing in C++. It's definitely not worth the effort for maybe like 1% more performance.

The reason I'm fairly sure about this is because I profiled the python code: 30 of the 40 seconds of runtime on a batch of 15 videos were spent in decoder.DecodeSingleSurface() which is pretty much only a high-level wrapper around a few CUDA / CUVID API calls. Most of the work seems to already be happening on the GPU, mostly bottlenecked by memcpy speed as far as I can tell.