清晰 · QingXi
Run the bundled Python script; do not recreate the processing workflow or load its source for routine use. Requires Windows x64, Python 3.11+, Pillow and a working Vulkan GPU. Use an existing verified Python environment. Read README.md only for setup/errors; report missing requirements instead of building an installer or switching to cloud/CPU processing.
Run
Resolve the skill directory and invoke:
python "<skill-dir>/scripts/qingxi.py" "<absolute-image-or-folder>" --kind photo
python "<skill-dir>/scripts/qingxi.py" "<image>" --kind illustration --preset ultra --format png
- Default HD is 2x width/height; Ultra is 4x. Both use realesrgan-x4plus; HD downsamples the 4x result. No extra sharpening or face restoration.
- Inspect the image to choose
--kind photoorillustration. Photo output is JPEG quality 95, 4:4:4; illustration/alpha output is PNG. Script auto mode uses file extensions, not visual classification. - PNG is lossless encoding, not unchanged AI pixels. Preserve composition and overall color; do not promise authentic recovery of unreadable details.
--output-dirselects destination. Otherwise outputs go beside inputs in 清晰_QingXi. Originals and existing results are never overwritten; folders are top-level snapshots, excluding current/legacy outputs.- First uncached run downloads the fixed official engine/model ZIP and checks its hash. Subsequent runs reuse the cache.
--offlineforbids downloads;--archiveimports the specified official ZIP. Never remove model caches during cleanup. - Support 8-bit JPG/PNG/WebP/BMP and single-page RGB/gray TIFF. Reject RAW, CMYK, high-bit-depth and multi-frame input until conversion is agreed.
- Preserve applicable ICC/EXIF, apply orientation and write 300 DPI metadata. Alpha is interpolated, not AI-generated. Large images use overlapping blocks and resource preflight; do not silently resize input.
Deliver
The script checks hashes, dimensions and output decoding, records failures and creates a local comparison for the first successful image. View it for text deformation, smoothing, false texture and halos. Keep successful batch outputs; do not rerun them without cause. Return output/comparison links, actual dimensions and relevant caveats. Never claim a clarity multiplier or that 300 DPI guarantees print quality.
