image-cooker 0.1.0

CLI that prepares web-optimized images: downscales then converts to WebP.

image-cooker

Bake your blog photos. Downscale once, re-encode as WebP, ship to the web.

image-cooker is a small, opinionated Python CLI that takes the giant JPEG/PNG files your camera produces and turns them into web-optimised WebP images that don't sink your Lighthouse score.

It does one thing — and it does it well: shrink, then re-encode.

---

Why

Modern cameras output 4000×6000-pixel JPEGs at 5–15 MB each. Browsers ignore the DPI metadata embedded in those files; the only things that matter for web performance are the pixel dimensions and the codec efficiency. So image-cooker only attacks those two levers:

1. Downscale so the longest edge is at most 2560 px (configurable). Aspect ratio is preserved; smaller images are never upscaled.
2. Re-encode as WebP at quality 85 (configurable), with EXIF stripped to keep files small and avoid leaking camera metadata.

A real-world photo blog post dropped from 140 MB to 47 MB (-66%) with no perceptible quality loss, just by running the defaults.

---

Features

• Single binary CLI powered by Typer — rich --help, validated input, sensible exit codes.
• JPEG and PNG in, WebP out. Alpha channel preserved automatically for transparent PNGs.
• Mirrors directory trees. Point it at ~/Pictures/blog/ and get back the same structure under your target folder, every file as .webp.
• Single-file mode when you just need to cook one photo.
• Recursive or shallow. Toggle with --recursive / --no-recursive.
• Lanczos resampling for the downscale step — the same algorithm professional image tools use.
• Per-file resilience. A bad file is logged and skipped; the rest of the batch keeps going. The process exits non-zero if anything failed.
• Zero hidden state. No config files, no caches, no surprise network calls. What you pass on the command line is what you get.

---

Quick start

uv run image-cooker --source ~/Pictures/blog --target ./optimised

Output:

~/Pictures/blog/IMG_0001.jpg -> ./optimised/IMG_0001.webp  3888x5175 -> 1923x2560  4.2 MB -> 1.4 MB (-66.7%)
~/Pictures/blog/IMG_0002.jpg -> ./optimised/IMG_0002.webp  6000x4000 -> 2560x1707  6.1 MB -> 2.0 MB (-67.2%)
...

42 converted, 0 failed. Total: 140.0 MB -> 47.0 MB (-66.4%)

---

Installation

From source (current)

git clone https://github.com/vsevolodbazhan/image-cooker.git
cd image-cooker
uv sync

Then run via uv run image-cooker ... or activate the venv and use the image-cooker console script directly.

Requirements

• Python 3.10+
• uv (recommended), or any PEP 517–compatible installer (pip, pipx, hatch, …)

---

Usage

Usage: image-cooker [OPTIONS]

  Prepare web-optimized WebP images from JPEG/PNG sources.

Option	Default	Description
--source PATH	(required)	Source file or directory containing JPEG/PNG images.
--target PATH	(required)	Output file (when source is a file) or directory (when source is a directory).
--max-edge INT	2560	Cap on the longest edge in pixels. Aspect ratio is preserved; smaller images are not upscaled.
--quality INT	85	WebP encoder quality, 1–100.
--recursive / --no-recursive	recursive	Recurse into subdirectories when source is a directory.
-h, --help		Show the help message and exit.

Examples

Optimise a whole blog photo folder, recursively:

uv run image-cooker --source ~/Pictures/blog --target ~/Sites/cooked

Cook a single hero image with custom settings:

uv run image-cooker \
  --source hero.jpg \
  --target hero.webp \
  --max-edge 1920 \
  --quality 88

Build a smaller thumbnail set in a sibling directory:

uv run image-cooker \
  --source ~/Pictures/blog \
  --target ./thumbs \
  --max-edge 960 \
  --quality 75

Process only the top level (no recursion):

uv run image-cooker --source ./photos --target ./out --no-recursive

---

How it works

The pipeline is deliberately small and easy to reason about:

1. Discover — walk --source (recursive by default), collect *.jpg, *.jpeg, *.png (case-insensitive). Non-image files are ignored.
2. Plan output paths — mirror the source tree under --target, replacing each file's extension with .webp. In single-file mode, --target is treated as the output file path.
3. Resize — if the longest edge exceeds --max-edge, downscale with Lanczos resampling and the aspect ratio preserved. Smaller images pass through untouched.
4. Normalise mode for the WebP encoder:
   • Opaque sources (JPEG, opaque PNG) → RGB
   • Transparent sources (RGBA PNG, palette+transparency) → RGBA
5. Encode with format="WEBP", quality=<q>, method=6. EXIF is intentionally not carried over.
6. Report — one line per file plus a final summary.

That's it. There is no caching, no parallelism, no hidden config — the code is short enough to read in a sitting.

---

Tuning guide

The defaults (--max-edge 2560, --quality 85) target high-quality photo blogs on retina displays. If you want to push further:

• Smaller files, same look: drop --quality to 82 or 80. Below ~75 you'll start seeing artefacts on smooth gradients (skies, skin).
• Smaller files for thumbnails or grid layouts: drop --max-edge. 1920 is fine for most blogs; 1280 is great for non-retina layouts.
• No-compromise archive copy: raise --quality to 92 and --max-edge to 3200. Files get noticeably bigger but quality is effectively transparent.

---

Development

git clone https://github.com/vsevolodbazhan/image-cooker.git
cd image-cooker
uv sync
uv run pre-commit install   # one-time: wire up commit + push hooks
uv run pytest

Quality gates run automatically:

• On git commit — ruff check --fix, ruff format, pyright
• On git push — pytest

To run them manually: uv run pre-commit run --all-files.

Project layout:

image-cooker/
├── pyproject.toml              # PEP 621 + PEP 735 dep groups
├── uv.lock                     # reproducible env
├── src/image_cooker/
│   ├── __init__.py
│   ├── __main__.py             # python -m image_cooker
│   ├── cli.py                  # Typer command, orchestration
│   └── processor.py            # discovery + image transform
└── tests/
    └── test_processor.py       # pipeline unit tests

The whole project is intentionally tiny — fewer than 250 lines of Python — and fully covered by pytest. Contributions welcome.

---

License

MIT

---

Acknowledgements

• Pillow for the imaging primitives
• Typer for the delightful CLI
• uv for fast, reproducible Python envs