Can I connect GPU frames to another interface without copying?

[caffeinism]

It would be great if everything could be done in ffmpeg, but I discovered that overlay_cuda doesn't support transparent frames (in fact, I realized there are no frames with alpha channels supported on the GPU to begin with). I want to decode video using nvdec (cuvid) in pyav, copy it to PyTorch for processing, then move it back to pyav for encoding with nvenc.

However, I see no signs that anyone has attempted this, so I'm wondering if it's actually possible. Is there anyone who might be able to help?

[caffeinism]

... Looking at the logs, it appears pyav immediately transfers decoded frames to the CPU and frees the original CUDA memory. Is it not possible to process?

[caffeinism]

#1893 (comment)