Compact Pretrained Models from the wiki below are used.
https://upscale.wiki/wiki/Model_Database#Real-ESRGAN_Compact_Custom_Models
This script will convert a video file using ten minutes worth of frames at a time to save disk space. The video fragments are then merged at the end of the process into a final video file.
Without a GPU this process would take weeks to process a 2 hour video.
Running test_gpu.py will list out Vulkan compatible GPUs and other info.
Make sure you are passing in a ffmpeg encoder that is compatible with your system.
The default is "av1_qsv" since I'm using an Intel Arc A750 GPU which supports AV1 hardware encoding.
Also make sure you have the latest Mesa drivers installed if on linux (see Notes below).
List of ffmpeg encoders is here: https://ffmpeg.org/ffmpeg-codecs.html#Video-Encoders.
Python 3.7 through 3.11.
Download and install ffmpeg: https://ffmpeg.org/download.html
pip install ncnn_vulkan
pip install wakepy>=0.7.2
pip install pillow
git clone https://github.com/davlee1972/upscale_video.git
Step 1 - Calibrate your GPU(s)
Run python test_gpus.py which will show you what gpus are available.
Run python test_gpus.py -g 0 -s 2 to test gpu 0 running x2 scaling.
Run python test_gpus.py -g 0,0 -s 2 to test gpu 0 using two workers in parallel.
Run python test_gpus.py -g 0,0,0 -s 2 to test gpu 0 using three workers in parallel.
Review timing logs to see how many workers your GPU can support before performance starts to degrade.
If you have a second, third, etc. gpu, repeat the process for the other gpus.
Run python test_gpus.py -g 1 -s 2 to test gpu 1 running x2 scaling.
Run python test_gpus.py -g 1,1 -s 2 to test gpu 1 using two workers in parallel.
Run python test_gpus.py -g 1,1,1 -s 2 to test gpu 1 using three workers in parallel.
Test your final GPU(s) configuration..
Run python test_gpus.py -g 0,0,0,0,1,1 -s 2 to test gpu 0 with 4 workers and gpu 1 with 2 workers in parallel.
-g, --gpus can be passed into upscale_video.py. If omitted upscale_video.py will just default to -g 0.
Step 2 - Sample your video upscaling if you want to apply any model filters.
This is a nice to have if you have older videos and want to remove grain, etc..
You can skip this step if you just want to apply AI upscaling only.
Run upscale_video.py with the -x, --extract_only option.
This will stop processing after frames extraction.
Run test_images.py on some extracted png files to sample what denoise level '-m n=???' to apply if needed. '-m a' can also be passed in to apply anime enhancements.
Run upscale_video.py with -r, --resume_processing to continue with upscaling.
Usage: python upscale_video.py -i INPUT_FILE -f FFMPEG
options:
-h, --help Show this help message and exit
-i, --input_file INPUT_FILE Input video file.
-o, --output_file OUTPUT_FILE Optional output video file location.
Default is input_file + ('.2x.' or '.4x.')
-f, --ffmpeg FFMPEG Location of ffmpeg.
-e, --ffmpeg_encoder FFMPEG_ENCODER ffmpeg encoder for video file. Default is av1_qsv.
-p, --pix_fmt PIX_FMT Pixel format used when merging frames into a video file.
Default is p010le which is 10 bit color.
-m, --models MODELS '-m a' will processing for anime to remove grain and color bleeding.
'-m n=???' will add processing to remove film grain.
Denoise level 1 to 30. 3 = light / 10 = heavy, etc.
'-m r' will add processing for real life imaging
Can include combinations like '-m a,n=3,r'
-s, --scale SCALE Scale 2 or 4. Default is 2. If using real life imaging (4x model),
scale will autoset to 4.
-t, --temp_dir TEMP_DIR Temp directory. Default is tempfile.gettempdir().
-g, --gpus GPUS Optional gpus to use. Example 0,1,1,2. Default is 0.
-b, --batch_size BATCH_SIZE Number of minutes to upscale per batch. Default is 10.
-r, --resume_processing Does not purge any data in temp_dir when restarting.
-x, --extract_only Exits after frames extraction. You may want to
run test_image.py on some extracted png files
to sample what denoise level to apply if needed.
Rerun with -r / --resume_processing to restart.
-l, --log_level LOG_LEVEL Logging level. logging.INFO is default
-d, --log_dir LOG_DIR Logging directory. logging directory
Usage: python test_images.py -i input_frames -o output_dir
options:
-h, --help Show this help message and exit
-i, --input_frames INPUT_FRAMES List of input frames in format like 1,3,5-7,10-12,15
-t, --temp_dir TEMP_DIR Temp directory where extracted frames are saved. Default is tempfile.gettempdir().
-o, --output_dir OUTPUT_DIR Output directory where test images will be saved.
-s, --scale SCALE Scale 2 or 4. Default is 2. If using real life imaging (4x model),
scale will autoset to 4.
-m, --models MODELS '-m a' will add processing for anime to remove grain and color bleeding.
'-m n=???' will add processing to remove film grain.
Denoise level 1 to 30. 3 = light / 10 = heavy, etc.
'-m r' will add processing for real life imaging
Can include combinations like '-m a,n=3,r'
-g, --gpus GPUS Optional gpus to use. Example 0,1,1,2. Default is 0.
Usage: python test_gpus.py -g gpus
options:
-h, --help Show this help message and exit
-g, --gpus GPUS Optional gpus to test. examples: 0 or 0,1 or 0,1,1 (to test same gpu twice)
-s, --scale SCALE Scale 2 or 4. Default is 2.
-r, --runs RUNS Number of test runs. Default is 10.
This python code is used to scale my existing 2k bluray collecion to 4k
Troubleshooting Ubuntu. It took me a while to figure out why my AMD gpu on my linux server wasn't working with Vulkan. Apparently you have to be a member of the ‘render’ group while running in a shell session.
In order to access GPU capabilities, a user needs to have the correct permissions on the system. The following will list the group assigned ownership of the render nodes, and list the groups the active user is a member of:
stat -c "%G" /dev/dri/render*
groups ${USER}If a group is listed for the render node which isn’t listed for the user, you will need to add the user to the group using gpasswd. In the following, the active user will be added to the ‘render’ group and a new shell spawned with that group active:
sudo gpasswd -a ${USER} render
newgrp render