python-ffmpeg | A python binding for FFmpeg which provides sync and async APIs | Video Utils library

I'll answer my own question, as I've been able to solve the issue with the help of kesh in the comments.

There are basically two things:

• vsync passthrough is required for the input video, to keep the number of frames
• another external tool (MKVToolNix) has to be used twice to extract timestamps from the initial video and apply them to the output

Below is the relevant code to perform the whole operation using python and subprocess. You can use the following line on both input and output video to check that the timestamps are indeed the same for each frame: ffprobe -show_entries packet=pts_time,duration_time,stream_index video.mkv

If anybody else is curious, I was able to answer my own question running a longer experiment.

(Presumably) the PCM encoder/decoder (used pcm_f32le for both) initially buffers its input excessively and the max buffer size appears to be sampling rate dependent. It maxes out at somewhere between 51200 - 52224.

Once the configuration log is posted on stderr, the output floodgate opens and eventually stabilizes to the expected number of output samples per input sample.

Here is a log of repeatedly writing 1024 samples at a time. The filter is afade so we expect the same number of output samples. The stdout read operation output is read 1024 bytes at a time to a queue on the reader thread and the main thread retrieves blocks with the timeout set to 10 ms.

Getting the same hash requires when writing and reading a video file requires careful attention.

Before comparing the hash, try to look at the video first.

Executing your code gave me the following output (first frame of video_2):

When the input (first frame of video) is:

I suggest the following modifications:

• Use AVI container (instead of MKV) for storing test_2 video in raw video format.
  AVI video container is originally designed for storing raw video.
  There could be a way for storing raw, or lossless RGB video in MKV container, but I am not aware of such option.
• Set the input pixel format of test_2 video.
  Add an argument: pixel_format='rgb24'.
  Note: I modified it to pixel_format='bgr24', because AVI supports bgr24 and not rgb24.
• Select video a lossless codec for test_2 video.
  You may select vcodec='rawvideo' (rawvideo codec is supported by AVI but not supported by MKV).

Note:
For getting equal hash, you need to look for lossless video codec that supports rgb24 (or bgr24) pixel format.
Most of the lossless codecs, converts the pixel format from RGB to YUV.
The RGB to YUV conversion has rounding errors that prevents equal hash.
(I suppose there are ways to get around it, but it's a bit complicated).

Here is your complete code with few modifications:

“Concatenate” means making one stream run after the other, but you want to merge both streams at the same time. So, remove the ffmpeg.concat step, and just pass both streams into one call to ffmpeg.output: