GPU-accelerated AI image upscaler.
sharp <input image folder path> <output image folder path> [<options>]
or
sharp_uv <input image folder path> <output image folder path> [<options>]
{f} : input relative file path
{d} : input relative directory path
{n} : input file name
{s} : input file stem
{e} : input file extension
{r} : output scaling ratio
{w} : output width
{h} : output height
--include <input_image_file_path_filter>
--exclude <input_image_file_path_filter>
--template <output_image_file_path_template={d}{s}.png>
--model <model_name=remacri>
--min-width <minimum_width=0>
--min-height <minimum_height=0>
--max-ratio <maximum_ratio=4>
--max-width <maximum_width=0>
--max-height <maximum_height=0>
--max-upscaled-width <maximum_upscaled_width=0>
--max-upscaled-height <maximum_upscaled_height=0>
--tile-size <tile_size=400|0>
--compression <compression>
--avif-compression <avif_compression=85>
--jpeg-compression <jpeg_compression=85>
--webp-compression <webp_compression=85>
--alpha-mode <alpha_mode=lanczos|realesrnet|remove>
--alpha-color <alpha_color=FFFFFF>
--skip
--cpu
--cuda
--rocm
--include and --exclude filters are applied in order, so a later --include can re-include paths excluded earlier.
If none of --cpu, --cuda, or --rocm is passed, Sharp uses CUDA when available, otherwise ROCm when available, otherwise CPU.
bsrgan— real-world blind super-resolutionbsrnet— smooth real-world blind super-resolutionhighfidelity— anime with high-frequency detaillsdirplus— LSDIR-trained photo upscaler; handles JPG compression and light noise/blurlsdirplusc— LSDIR-trained RealESRGAN finetune; strong JPG compression removalnomos8ksc— realistic photo upscaler for JPG compression and slight blurrealanime— anime and mangarealdigital— digital artrealesrgan— sharp texturesrealesrgeneral— fast lightweight general-purpose upscalerrealesrnet— smooth output with minimal invented detailremacri— strong on skin, faces, and fine texturesultramix— balanced detail and smoothnessultrasharp— aggressive detail recoveryuniscalerestore— strong compression/noise removal; alternative to BSRGAN
sharp "IN/" "OUT/"
sharp "IN/" "OUT/" --model remacri
sharp "IN/" "OUT/" --include "**/*.jpg" --include "**/*.png" --template "{d}{s}.jpg" --skip
sharp "IN/" "OUT/" --include "**/*.jpg" --include "**/*.png" --template "{d}{s}_{r}x.jpg" --max-ratio 3 --skip
sharp "IN/" "OUT/" --include "**/*.jpg" --include "**/*.png" --template "{d}{s}_{w}x{h}.png" --max-width 3840 --compression 85 --skip
sharp "IN/" "OUT/" --include "**/*.jpg" --include "**/*.png" --exclude "alpha/*.*" --exclude "anime/*.*" --exclude "architectural/*.*" --exclude "digital/*.*" --exclude "**/small/*.*" --include "building/small/*.*" --template "{d}{s}.jpg" --min-width 3840 --max-upscaled-width 3840 --skip
Run one of:
install_packages_cuda.bat/install_uv_packages_cuda.bat— NVIDIA CUDAinstall_packages_rocm.bat/install_uv_packages_rocm.bat— AMD ROCminstall_packages_cpu.bat/install_uv_packages_cpu.bat— CPU only
- Python 3.12.10
- CUDA 12.4 (NVIDIA) or ROCm 7.2.1 (AMD), optional
- Only reads and writes AVIF, JPEG, PNG, and WebP images.
- Output resolution is limited to 4× the original resolution.
0.1
Eric Pelzer (ecstatic.coder@gmail.com).
This project is licensed under the GNU General Public License version 3.
See the LICENSE.md file for details.
