av1-scd

Command line tool to quickly detect scene change and generate config file for encoder to force keyframe for Encoding video.

• Installation
• Encoder format
• Scene detection method
• Dependencies
• Checking Keyframe of video
• Usage
• Parameter

Support encoder format

• x264
• x265
• SvtAv1EncApp
• ffmpeg
• av1an
• xav

Support scene detection method

• Pyscenedetect
• WWXD and Scxvid (vapoursynth)
• av-scenechange
• ffmpeg (using scene score and scdet filter)
• TransNetV2 This project use onnxruntime and opencv-python not tensorflow in the original project

Dependencies

This is not require dependencies for a full list of dependencies checkout pyproject.toml
if any dependencies is missing it will error out anyway.

• Pymediainfo
• Pyscenedetect
• vapoursynth
• ffmpeg
• av-scenechange
• opencv-python
• numpy
• tqdm
• onnxruntime
• TransetV2 The tensorflow model is in "inference/transnetv2-weights/model" folder of the original project if you want to use with this program you need onnx model which you can download it here
• pyav_scenechange optional dependepcies for av-scenechange

Checking Keyframe of video

1. Use LosslessCut to check
2. FFprobe command : The command list keyframe of video

• Bash (linux)

input="input.mkv"

# Get frame rate as decimal
fps=$(ffprobe -v 0 -select_streams v:0 -show_entries stream=r_frame_rate \
      -of default=nokey=1:noprint_wrappers=1 "$input" | awk -F'/' '{printf "%.0f", $1 / ($2 ? $2 : 1)}')

# Extract keyframe PTS and convert to frame number
ffprobe -loglevel error -select_streams v:0 \
  -show_entries packet=pts_time,flags -of csv=print_section=0 "$input" |
awk -F',' -v fps="$fps" '/K/ {printf "%.0f\n", $1 * fps}'

The report keyframe may differ slightly (usually 1,2 or 3 frames) depend on program (This is normal)

Installtion

1. From Pypi

   pipx install av1-scd # install with minimal dependencies
   pipx install av1-scd[vsxvid,transnet] # install with optional vsxvid and transnet dependencies. All option are [vsxvid, transnet, pyscene]
   

2. Build wheel

   git clone https://github.com/Khaoklong51/av1-scd.git
   cd av1-scd
   python -m build --wheel # or 'uv build' if you have uv.
   pipx install dist/*.whl # install with minimal dependencies
   pipx install dist/*.whl[vsxvid,transnet] # install with optional vsxvid and transnet dependencies. All option are [vsxvid, transnet, pyscene]
   

3. Wheel files

   Wheel file can be download at release section
   
   then install with pipx or your prefered package manager

   pipx install av1-scd.whl # install with minimal dependencies
   pipx install av1-scd.whl[vsxvid,transnet] # install with optional vsxvid and transnet dependencies. All option are [vsxvid, transnet, pyscene]
   

Usage

av1-scd -i input.mp4 -o x265.cfg -f x265

Parameter

usage: av1-scd.exe [-h] -i INPUT [-o OUTPUT] [--min-scene-len MIN_SCENE_LEN]
                   [--max-scene-len MAX_SCENE_LEN]
                   [--scd-method {pyscene,vsxvid,av-scenechange,ffmpeg-scene,ffmpeg-scdet,transnetv2}]
                   [--track TRACK] -f {x264,x265,svt-av1,av1an,ffmpeg,xav}
                   [--print] [--log-level {debug,info,warning,error}]
                   [--threshold THRESHOLD] [--ignore-scene-len] [--version]
                   [--pysc-decode {opencv,pyav,moviepy}]
                   [--pysc-method {adaptive,content,threshold,hash,histogram}]
                   [--pysc-downscale PYSC_DOWNSCALE]
                   [--vs-source {bestsource,ffms2,lsmash}]
                   [--vsxvid-height VSXVID_HEIGHT]
                   [--transnet-model TRANSNET_MODEL]
                   [--ffmpeg-filter FFMPEG_FILTER]
                   [--avsc-analyze-speed {standard,fast}]
                   [--avsc-detect-flashes | --no-avsc-detect-flashes]
                   [--avsc-lookahead AVSC_LOOKAHEAD]
                   [--avsc-score-mode {none,xav}]
                   [--avsc-mode {binary,pyav_scenechange}]
                   [--avsc-width AVSC_WIDTH] [--avsc-height AVSC_HEIGHT]

py-video-encode v3.0.1

options:
  -h, --help            show this help message and exit
  -i, --input INPUT     Path to input file
  -o, --output OUTPUT   Path to output file
  --min-scene-len MIN_SCENE_LEN
                        min lenght for scene detection | Default is 1 sec of
                        video
  --max-scene-len MAX_SCENE_LEN
                        max lenght for scene detection | Default is 10 sec of
                        video
  --scd-method {pyscene,vsxvid,av-scenechange,ffmpeg-scene,ffmpeg-scdet,transnetv2}
                        scene detection method | Default is pyscene
  --track TRACK         Track number for video (Index start at 1) | Default is
                        1
  -f, --format {x264,x265,svt-av1,av1an,ffmpeg,xav}
                        format of keyframe to feed program
  --print               print data to stdout. this will disable the last
                        helper massage
  --log-level {debug,info,warning,error}
                        log level output to console | Default is info
  --threshold THRESHOLD
                        treshold for scene change. Only affect pyscene and
                        ffmpeg scd based method
  --ignore-scene-len    skip keyframe processing that make scene len lenght
                        exactly follow the value if method does not expose way
                        to set min or max value of scene. This is not the same
                        as 1 min and 9999 max scene len
  --version             print version

pyscene:
  Extra option for pyscene scene detection method

  --pysc-decode {opencv,pyav,moviepy}
                        Decode method for pyscene detect | Default is opencv
  --pysc-method {adaptive,content,threshold,hash,histogram}
                        Scene detect method for pyscene detect | Default is
                        adaptive
  --pysc-downscale PYSC_DOWNSCALE
                        Downscale factor for pyscene detect method, can be
                        either auto or number(int). To disable set this to 1 |
                        Default is auto

vapoursynth:
  Extra option for vapousynth to perform vs-xvid scene detection method

  --vs-source {bestsource,ffms2,lsmash}
                        Source method for vapoursynth | Default is ffms2
  --vsxvid-height VSXVID_HEIGHT
                        Height for vsxvid processing | Default is video height

transnet:
  Extra option for transnetv2 model

  --transnet-model TRANSNET_MODEL
                        Path to onnx transet model

av-scenechnage:
  Extra option for av-scenechange

  --ffmpeg-filter FFMPEG_FILTER
                        Extra option to go in to -filter:v in ffmpeg for
                        piping. Useful for downscaling video if select the
                        binary mode
  --avsc-analyze-speed {standard,fast}
                        Set av-scenechange analysis speed | Default is
                        standard
  --avsc-detect-flashes, --no-avsc-detect-flashes
                        Enable or disable av-scenechange flash detection
  --avsc-lookahead AVSC_LOOKAHEAD
                        Set custom av-scenechange lookahead distance the value
                        must be an integer more than 0 | Default is 5
  --avsc-score-mode {none,xav}
                        Scenechange calculation mode if select xav score will
                        using weight score logic (xav score mode is only
                        support if you select the python binding mode) |
                        Default is none
  --avsc-mode {binary,pyav_scenechange}
                        Select whether to use native python binding or binary
                        | Default is binary
  --avsc-width AVSC_WIDTH
                        Video width use to detect scenechange. Useful for
                        downscaling video
  --avsc-height AVSC_HEIGHT
                        Video height use to detect scenechange. Useful for
                        downscaling video