The following line executes the CLI command.
python -m waifu2x.cli -h
options:
-h, --help show this help message and exit
--model-dir MODEL_DIR
model dir (default: None)
--noise-level {0,1,2,3}, -n {0,1,2,3}
noise level (default: 0)
--method {scale4x,scale2x,noise_scale4x,noise_scale2x,scale,noise,noise_scale}, -m {scale4x,scale2x,noise_scale4x,noise_scale2x,scale,noise,noise_scale}
method (default: noise_scale)
--gpu GPU [GPU ...], -g GPU [GPU ...]
GPU device ids. -1 for CPU (default: [0])
--batch-size BATCH_SIZE
minibatch_size (default: 4)
--tile-size TILE_SIZE
tile size for tiled render (default: 256)
--output OUTPUT, -o OUTPUT
output file or directory (default: None)
--input INPUT, -i INPUT
input file or directory. (*.txt, *.csv) for image list (default: None)
--tta use TTA mode (default: False)
--disable-amp disable AMP for some special reason (default: False)
--image-lib {pil,wand}
image library to encode/decode images (default: pil)
--depth DEPTH bit-depth of output image. enabled only with `--image-lib wand` (default: None)
--format {png,webp,jpeg}, -f {png,webp,jpeg}
output image format (default: png)
--style {art,photo,scan,art_scan}
style for default model (art/scan/photo). Ignored when --model-dir option is specified. (default: None)
--grayscale Convert to grayscale format (default: False)
--recursive, -r process all subdirectories (default: False)
--resume skip processing when output file is already exist (default: False)
--max-fps MAX_FPS max framerate. output fps = min(fps, --max-fps) (video only) (default: 128)
--crf CRF constant quality value. smaller value is higher quality (video only) (default: 20)
--preset {ultrafast,superfast,veryfast,faster,fast,medium,slow,slower,veryslow,placebo}
encoder preset option (video only) (default: ultrafast)
--tune {film,animation,grain,stillimage,fastdecode,zerolatency} [{film,animation,grain,stillimage,fastdecode,zerolatency} ...]
encoder tunings option (video only) (default: ['zerolatency'])
--yes, -y overwrite output files (video only) (default: False)
--rotate-left Rotate 90 degrees to the left(counterclockwise) (video only) (default: False)
--rotate-right Rotate 90 degrees to the right(clockwise) (video only) (default: False)
--vf VF video filter options for ffmpeg. (video only) (default: )
--grain add noise after denosing (video only) (default: False)
--grain-strength GRAIN_STRENGTH
noise strength (video only) (default: 0.05)
--pix-fmt {yuv420p,yuv444p}
pixel format (video only) (default: yuv420p)
When DEBUG environment variable is defined, DEBUG level log will be printed.
DEBUG=1 python -m waifu2x.cli -h
Denoise level 0 (-n noise_level)
python -m waifu2x.cli -m noise -n 0 -i tmp/images -o tmp/out
2x
python -m waifu2x.cli -m scale -i tmp/images -o tmp/out
2x, webp output
python -m waifu2x.cli -m scale -i tmp/images -o tmp/out -f webp
python -m waifu2x.cli -m scale -i tmp/images/image.jpg -o tmp/out/image.webp
2x + Denoise level 3
python -m waifu2x.cli -m noise_scale -n 3 -i tmp/images -o tmp/out
4x + Denoise level 3
python -m waifu2x.cli -m noise_scale4x -n 3 -i tmp/images -o tmp/out
Using photo model (--style photo)
python -m waifu2x.cli --style photo -m noise_scale4x -n 3 -i tmp/images -o tmp/out
Also, for photo models, larger --tile-size will give better results (less tile seam/border artifact)
python -m waifu2x.cli --style photo -m noise_scale4x -n 3 --tile-size 640 --batch-size 1 -i tmp/images -o tmp/out
Using art_scan model (--style scan or --style art_scan)
python -m waifu2x.cli --style scan -m noise_scale4x -n 3 -i tmp/images -o tmp/out
With model dir
python -m waifu2x.cli --model-dir ./waifu2x/pretrained_models/upconv_7/photo/ -m noise_scale -n 1 -i tmp/images -o tmp/out
With multi GPU (--gpu gpu_ids)
python -m waifu2x.cli --gpu 0 1 2 3 -m scale -i tmp/images -o tmp/out
With TTA
python -m waifu2x.cli --tta -m scale -i tmp/images -o tmp/out
Note that TTA is not valid for photo and art_scan, which are GAN-based models.
With TTA, increase mini-batch size
python -m waifu2x.cli --tta --batch-size 16 -m scale -i tmp/images -o tmp/out
First of all, the waifu2x models are designed for images, not for video stream.
examples
python -m waifu2x.cli --style art --method noise_scale -i tmp/input_video.mp4 -o tmp/output_dir
python -m waifu2x.cli --style art_scan --method noise -n 3 -i tmp/input_video.mp4 -o tmp/output_video.mp4
high quality encoding
python -m waifu2x.cli --crf 16 --preset medium --pix-fmt yuv444p --style art_scan --method noise -n 3 -i tmp/input_video.mp4 -o tmp/output_video.mp4
drop to 30fps from 60fps video (fps = min(original fps, --max-fps).
python -m waifu2x.cli --max-fps 30 --style art_scan --method noise -n 3 -i tmp/input_video.mp4 -o tmp/output_video.mp4
1fps video (for preview)
python -m waifu2x.cli --max-fps 1 --style art_scan --method noise -n 3 -i tmp/input_video.mp4 -o tmp/output_video.mp4
fix rotation (width height swap, --rotate-left(counterclockwise) or --rotate-right(clockwise))
python -m waifu2x.cli --rotate-left --style photo --method noise -n 3 -i tmp/input_video.mp4 -o tmp/output_video.mp4
deinterlace input video stream. (you can use ffmpeg's video filter with --vf option)
python -m waifu2x.cli --vf yadif --style photo --method noise -n 3 -i tmp/input_video.mp4 -o tmp/output_video.mp4
add noise after denosing
python -m waifu2x.cli --grain --style photo --method noise -n 3 -i tmp/input_video.mp4 -o tmp/output_video.mp4
use lightweight(older) models
cunet/art is best trade-off between performance and quality
python -m waifu2x.cli --model-dir ./waifu2x/pretrained_models/cunet/art --method noise -n 3 -i tmp/input_video.mp4 -o tmp/output_video.mp4
upconv_7/art and upconv_7/photo are the fastest, but lowest quality
python -m waifu2x.cli --model-dir ./waifu2x/pretrained_models/upconv_7/art --method noise -n 3 -i tmp/input_video.mp4 -o tmp/output_video.mp4
python -m waifu2x.cli --model-dir ./waifu2x/pretrained_models/upconv_7/photo --method noise -n 3 -i tmp/input_video.mp4 -o tmp/output_video.mp4