Image Upscaler

Use the bundled Python scripts to run Upscayl's Vulkan-based NCNN engine locally. Download models only when selected. Preserve the source aspect ratio by default and never overwrite an input file.

Workflow

1. Inspect the input path, file type, dimensions, and available disk space.
2. Inspect the image visually and recommend one profile. If the choice materially affects the result, give the user a compact choice with the advantage, disadvantage, and additional download size. Do not block if the user does not choose; use default.
3. Choose a profile:
   • default: balanced and conservative; use when content is mixed or uncertain. Model download: about 32 MB.
   • fast: smallest and fastest; weaker fine-detail reconstruction. Model download: about 2.5 MB.
   • photo: natural photographic texture; less crisp on text and architecture. Model download: about 32 MB.
   • portrait: use for compressed, noisy, or rough natural portraits that benefit from smoother rendering. It avoids face replacement but can remove fine skin texture. Use default for already-clean AI portraits and sharp studio headshots. Model download: about 32 MB, shared with photo.
   • digital-art: clean anime, illustration, and icon edges; may oversimplify photographic texture. Model download: about 8.6 MB.
   • sharp: strongest apparent edge detail; may exaggerate noise, halos, skin, or text artifacts. Model download: about 32 MB.
4. If the selected runtime or model is not cached, explain the download and disk cost, then run setup. Read platforms.md for compatibility and cache sizes.
5. Run the upscaler or compressor. Prefer PNG for maximum upscale quality; use JPEG for strict file-size targets.
6. Verify that the output exists, has the requested dimensions, and keeps the source ratio. Report the selected profile, output path, dimensions, and file size.

The agent performs the meaningful automatic selection after visually inspecting the image. CLI --profile auto uses filename hints only and otherwise falls back conservatively to default; it does not claim reliable face or scene classification without an extra vision package. Read algorithms.md when comparing profiles or discussing heavier optional algorithms.

For portraits, inspect source quality before selecting by subject alone: choose default for clean AI-generated or studio portraits; choose portrait only when a natural portrait is noisy, compressed, or visually rough. This rule is based on same-size comparison testing, where portrait produced smoother skin but removed detail from an already-clean AI portrait.

Prepare the Runtime

Run from this skill directory:

python scripts/setup.py --model high-fidelity

List choices, trade-offs, cache status, and incremental downloads without installing anything:

python scripts/setup.py --list-models

Prepare only the selected profile:

python scripts/setup.py --profile photo
python scripts/setup.py --profile digital-art
python scripts/setup.py --profile sharp

The downloader tries the local cache, configured mirrors, a verified GitHub acceleration endpoint, and the official source. Every downloaded artifact must match the pinned SHA-256 digest. The upscaler re-verifies the cached executable and selected model before every execution, including when IMAGE_UPSCALER_CACHE overrides the cache directory. To use an organization or domestic mirror, set IMAGE_UPSCALER_MIRRORS to one or more comma-separated URL prefixes before setup. Each prefix is prepended to the official URL. Example:

IMAGE_UPSCALER_MIRRORS="https://mirror.example/proxy/" python scripts/setup.py

Use --offline to forbid all network access. Use --platform windows-x64 --download-only to prefetch Windows files from another machine.

Upscale Images

Create a 4K image whose long edge is 3840 pixels while preserving the original ratio:

python scripts/upscale.py INPUT --output OUTPUT.png --target 4k

Useful variants:

python scripts/upscale.py INPUT --target 1k
python scripts/upscale.py INPUT --target 2k --profile fast
python scripts/upscale.py INPUT --target 4k --profile portrait
python scripts/upscale.py INPUT --target 4k --profile digital-art
python scripts/upscale.py INPUT --target 4k --profile sharp
python scripts/upscale.py INPUT --target 8k --profile photo --tile 256
python scripts/upscale.py INPUT --target scale4 --format png
python scripts/upscale.py INPUT_FOLDER --output OUTPUT_FOLDER --target 4k
python scripts/upscale.py INPUT --target 4k --format jpg --compression 92

The screen-oriented presets use long edges: 1k=1920, 2k=2560, 4k=3840, and 8k=7680. A 16:9 4K landscape becomes 3840×2160; portraits and squares retain their own ratios. scale4 keeps the model's native 4× output. Use --max-edge N for an exact custom long edge. Current local models are native 4× models; an 8K preset beyond the native result uses exact high-quality resizing and must not be described as recovered factual detail. For 8K, expect high memory and temporary-disk use; start with --tile 256 if automatic tiling fails.

Exact 2K/4K output supports PNG and JPEG without extra dependencies. WebP is supported with --target scale4; use PNG when an exact target and transparency both matter.

Compress to a Target Size

Compress a roughly 1MB image to no more than 200KiB:

python scripts/compress.py INPUT --output OUTPUT.jpg --target-kb 200
python scripts/compress.py INPUT --target-mb 1.5 --keep-dimensions
python scripts/compress.py INPUT_FOLDER --output OUTPUT_FOLDER --target-kb 200 --recursive

The compressor first finds the highest JPEG quality that fits. If quality alone is insufficient, it reduces dimensions proportionally while respecting --min-edge (default 640px). It never stretches or overwrites the source. JPEG does not preserve transparency; convert the background explicitly before compression when its color matters.

Use --keep-dimensions when reducing resolution is forbidden. Existing output files receive a -v2, -v3, and so on suffix unless --overwrite is explicitly supplied. Use --dry-run before a large folder job.

Inspect Before Processing

Inspect one file or a folder before choosing an upscale or compression target:

python scripts/inspect.py INPUT
python scripts/inspect.py INPUT_FOLDER --recursive --json

The report includes dimensions, simplified aspect ratio, megapixels, and file size.

Safety and Quality Rules

• Never delete or replace source images.
• Do not install GPU drivers, package managers, or system dependencies automatically.
• Do not bypass SHA-256 verification, even when using a mirror.
• Treat IMAGE_UPSCALER_CACHE and IMAGE_UPSCALER_MIRRORS as trusted configuration; execution-time hashes still reject modified runtime or model files.
• Do not upload user images; processing must stay local.
• Avoid TTA unless the user explicitly accepts a roughly 8× processing slowdown.
• On GPU-memory errors, retry with --tile 256, then --tile 128.
• If Vulkan is unavailable, stop and explain the platform limitation; do not silently switch to an unrelated cloud service.
• Treat visual enhancement as reconstruction, not recovery of factual detail. Do not claim that invented texture is original information.

References

• Read platforms.md for supported systems, storage, cache, mirror, and troubleshooting details.
• Read algorithms.md for profile selection, benefits, limitations, and optional heavy algorithms.
• Read licenses.md before redistributing the engine or model files.