Sharp

GPU-accelerated AI image upscaler.

Command line

sharp <input image folder path> <output image folder path> [<options>]

or

sharp_uv <input image folder path> <output image folder path> [<options>]

Template variables

{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

Options

--include <input_image_file_path_inclusion_filter>
--exclude <input_image_file_path_exclusion_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

If none of --cpu, --cuda, or --rocm is passed, Sharp uses CUDA when available, otherwise ROCm when available, otherwise CPU.

Models

• bsrgan — real-world blind super-resolution
• bsrnet — smooth real-world blind super-resolution
• highfidelity — anime with high-frequency detail
• realanime — anime and manga
• realdigital — digital art
• realesrgan — sharp textures
• realesrnet — smooth output with minimal invented detail
• remacri — strong on skin, faces, and fine textures
• ultramix — balanced detail and smoothness
• ultrasharp — aggressive detail recovery

Samples

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

Install

Run one of:

• install_packages_cuda.bat / install_uv_packages_cuda.bat — NVIDIA CUDA
• install_packages_rocm.bat / install_uv_packages_rocm.bat — AMD ROCm
• install_packages_cpu.bat / install_uv_packages_cpu.bat — CPU only

Dependencies

• Python 3.12.10
• CUDA 12.4 (NVIDIA) or ROCm 7.2.1 (AMD), optional

Limitations

• Only reads and writes AVIF, JPEG, PNG, and WebP images.
• Output resolution is limited to 4× the original resolution.

Version

0.1

Author

Eric Pelzer (ecstatic.coder@gmail.com).

License

This project is licensed under the GNU General Public License version 3.

See the LICENSE.md file for details.