thumbleweed 🌵

Unified image hashing, thumbnailing, and compression for Python.
Rust-powered via PyO3 + maturin. Zero mandatory dependencies.

• ✅ ThumbHash — compact image placeholder hashes (drop-in for thumbhash & fast-thumbhash)
• ✅ BlurHash — smooth gradient placeholders (drop-in for blurhash-python)
• ✅ ColorThief — dominant colour + palette extraction (drop-in for colorthief & fast-colorthief)
• ✅ Thumbnails — real rasterised thumbnails for images, videos (MP4), and PDFs, with a pure-Rust crude fallback that needs no ffmpeg, no pdfium
• ✅ Compression — optional pixo-powered JPEG/PNG re-encoder, auto-applied to thumbnail output (mozjpeg/oxipng-class quality, pure Rust)
• ✅ Python 3.11 – 3.15 (including free-threaded 3.13t / 3.14t / 3.15t)
• ✅ Pre-built wheels for Windows | Linux x86, AMD64, ARM64 and macOS Intel Based | Apple Silicon
• ✅ Pillow > 13.0 integration (optional)
• ✅ Typed (py.typed + .pyi stubs)
• ✅ Pure-Rust core — no C extensions, no NumPy required

Installation

pip install thumbleweed
# with Pillow helpers:
pip install "thumbleweed[pillow]"

Optional Rust-side features

A few capabilities are gated behind cargo features so the default wheel stays small and portable. The Python install extras are documentation hooks — to actually enable a feature you must build the wheel from source with the matching cargo feature:

Build from source with extras enabled:

# Compression only (lightweight, pure Rust)
MATURIN_PEP517_ARGS="--features pixo" pip install thumbleweed --no-binary thumbleweed

# Compression + full pdf/video backend (requires system pdfium + ffmpeg)
MATURIN_PEP517_ARGS="--features pixo,auto-thumbnail" \
    pip install thumbleweed --no-binary thumbleweed

Without these features the crude thumbnail engine is still available (MP4 cover-art atom scrape + PDF JPEG-stream scrape + deterministic gradient placeholder) and the compress module degrades to a safe pass-through.

Note that auto-thumbnail is currently an all-or-nothing upstream feature: enabling it pulls in both PDF and video support together rather than as separate cargo toggles.

Development with uv

curl -LsSf https://astral.sh/uv/install.sh | sh
make sync
make test

All import paths work:

import thumbleweed   # the unified package
import thumbhash     # ThumbHash only
import blurhash      # BlurHash only
import colorthief    # ColorThief
from thumbleweed import thumbnail   # rasterised thumbnails
from thumbleweed import compress    # optional pixo compression

Quick-start

ThumbHash

import thumbhash as th

# Encode — rgba_bytes must be bytes/bytearray of length w*h*4 (R G B A, non-premultiplied)
hash_bytes: bytes = th.encode(w, h, rgba_bytes)

# Decode
w_out, h_out, rgba_out = th.decode(hash_bytes)

# Helpers
r, g, b, a = th.average_rgba(hash_bytes)           # dominant colour [0, 1]
ratio = th.approximate_aspect_ratio(hash_bytes)     # width / height

BlurHash

import blurhash as bh

# Encode
hash_str: str = bh.encode(rgba_bytes, cx=4, cy=3, width=w, height=h)

# Decode
rgba: bytes = bh.decode(hash_str, width=64, height=64)

Pillow images / BytesIO

from PIL import Image
import io

# From a Pillow Image
import thumbhash as th
img = Image.open("photo.jpg")
hash_str = th.encode_image(img)             # base64 string; input is resized internally
placeholder = th.decode_image(hash_str)     # → RGBA Image, ≈32 px

# From a BytesIO object
buf = io.BytesIO(open("photo.jpg", "rb").read())
hash_str = th.encode_image(buf)

# BlurHash
import blurhash as bh
hash_str = bh.encode_image(img, cx=4, cy=3)
placeholder = bh.decode_image(hash_str, width=64, height=64)

ColorThief

import colorthief as ct

# From encoded image bytes (PNG, JPEG, WebP, …)
dominant = ct.get_color(image_bytes)                    # → (r, g, b)
palette = ct.get_palette(image_bytes, color_count=5)   # → [(r, g, b), (r, g, b), ...]

# From a file path
dominant = ct.get_color("photo.jpg")

# From a BytesIO object
import io
buf = io.BytesIO(open("photo.jpg", "rb").read())
dominant = ct.get_color(buf)                # accepts bytes, BytesIO, file path, or PIL Image
palette = ct.get_palette(buf, color_count=5)

# Class-based API (drop-in for the colorthief package)
thief = ct.ColorThief("photo.jpg")
dominant = thief.get_color()
palette = thief.get_palette(color_count=8)

thumbleweed (unified)

import thumbleweed

# ThumbHash raw-pixel API
hash_bytes = thumbleweed.thumbhash_encode(w, h, rgba)  # raw ThumbHash bytes
w, h, rgba = thumbleweed.thumbhash_decode(hash_bytes)

# ThumbHash image helper API
hash_str = thumbleweed.thumbhash_encode_image(image_bytes)  # base64 string like BlurHash

# BlurHash
hash_str = thumbleweed.blurhash_encode(rgba, 4, 3, w, h)
rgba = thumbleweed.blurhash_decode(hash_str, 64, 64)

# ColorThief
dominant = thumbleweed.colorthief_get_color(image_bytes)
palette = thumbleweed.colorthief_get_palette(image_bytes, 5, 10)

Thumbnails (images / videos / PDFs)

from thumbleweed import thumbnail

# Works on JPEG, PNG, WebP, GIF, BMP, MP4, MKV/WebM, PDF —
# accepts bytes, BytesIO, file paths, pathlib.Path, and PIL.Image.
jpeg_bytes = thumbnail.create("holiday.mp4", width=256, height=256)
png_bytes  = thumbnail.create("manual.pdf", format="png", width=128, height=128)
webp_bytes = thumbnail.create("photo.jpg", format="webp", width=64, height=64)

# Save directly
thumbnail.save("video.mp4", "video_thumb.jpg", width=256, height=256)

# Inspect the wheel's capabilities
print(thumbnail.detect_kind("file.pdf"))     # → "pdf"
print(thumbnail.available_backends())         # → ['crude'] or ['crude', 'auto-thumbnail']

Two engines ship in the wheel:

• crude (always available): pure-Rust pipeline. Resizes images via the image crate; scrapes embedded cover art from MP4 (covr atom) and the first JPEG (/DCTDecode) image stream from PDFs; otherwise generates a deterministic colour-gradient placeholder.
• auto-thumbnail (cargo feature): wraps the auto-thumbnail crate for proper PDF rasterisation (pdfium) and video frame extraction (ffmpeg).

Default engine="auto" picks auto-thumbnail when present, falling back to crude on any failure — you always get something back.

Compression (optional, pixo-backed)

from thumbleweed import compress, thumbnail

if compress.is_available():
    smaller = compress.compress(open("photo.jpg", "rb").read(), quality=85)
    bytes_saved = compress.compress_path("photo.jpg")  # in-place

# Auto-applied to thumbnail output (the default — pass compress=False to skip)
thumb = thumbnail.create("photo.jpg", width=256, height=256)            # compressed
thumb = thumbnail.create("photo.jpg", width=256, height=256, compress=False)

When the wheel is not built with --features pixo, every function in thumbleweed.compress becomes a no-op pass-through, so compress=True remains safe to leave on as the default everywhere.

Project structure

thumbleweed/
├── src/
│   ├── lib.rs            # PyO3 module — Python bindings
│   ├── thumbhash.rs      # Pure Rust ThumbHash encode/decode
│   ├── blurhash.rs       # Pure Rust BlurHash encode/decode
│   ├── colorthief.rs     # ColorThief — dominant colour & palette extraction
│   ├── thumbnail.rs      # Rasterised thumbnails (crude + auto-thumbnail)
│   └── compress.rs       # Optional pixo-powered JPEG/PNG re-encoder
├── python/
│   ├── thumbleweed/      # Main package — re-exports everything
│   │   ├── thumbnail.py  # High-level thumbnail submodule
│   │   └── compress.py   # High-level compression submodule
│   ├── thumbhash/        # Backward-compatible ThumbHash shim
│   ├── blurhash/         # BlurHash shim
│   └── colorthief/       # ColorThief shim
├── tests/
│   ├── test_thumbhash.py # 70+ ThumbHash tests
│   ├── test_blurhash.py  # 28+ BlurHash tests
│   ├── test_colorthief.py# ColorThief tests
│   ├── test_thumbnail.py # 153 thumbnail tests across all jpg / mp4 / pdf fixtures
│   ├── test_compress.py  # 60+ compression + integration tests
│   └── test_imports.py   # import / version-consistency tests
└── Cargo.toml

API reference

ThumbHash (import thumbhash)

BlurHash (import blurhash)

ColorThief (import colorthief)

Thumbnail (from thumbleweed import thumbnail)

Compress (from thumbleweed import compress)

Performance Benchmark Results

Benchmark configuration: 5 rounds × 100 iterations (pure-Python libraries use 5 iterations). Input corpus: all real image fixtures in tests/ (one.jpg, two.jpg, four.jpg, OPS.jpg). All times are mean per-call latency. Lower is better.

(The above values are updated by a benchmark script ocassionally)

Notes on ColorThief timing: thumbleweed includes image decode in its timing because it accepts raw encoded bytes from the real test fixtures, while fast-colorthief also reads from in-memory file-like objects in these benchmarks. This measures realistic end-to-end usage rather than just the inner palette routine.

Building from source

git clone https://github.com/New-Elysium/thumbleweed.git
cd thumbleweed
make sync   # sync uv environment + install editable
make test   # run Python + Rust tests

Make targets

• make sync — sync uv environment and install the extension in editable mode
• make test — run Python tests and Rust tests
• make dist — build wheels into dist/
• make upload — upload dist/* with twine

Running Benchmarks

Benchmarks require bench dependencies which are not compatible with every Python version (blurhash-python has no wheels for 3.13+, fast-colorthief none for 3.15+). To avoid conflicts, run them with a fully supported Python version (e.g. 3.12):

# Sync using a Python version that supports all benchmark dependencies
uv sync --group bench --python 3.12
uv run python tests/bench_comparison.py

# Refresh the benchmark table in this README
uv run python scripts/update_readme.py

Licence

MIT — see LICENCE.