Pillow guides and fixes

Resize images in Python without distortion: thumbnail(), ImageOps.contain, cover, fit and pad compared with real outputs, plus EXIF rotation, batch resizing and measured speed tips.

Draw text on images with Pillow's ImageDraw: load fonts, centre with anchors, wrap long text, add outlines and semi-transparent boxes, fit text to a box, emoji and non-Latin scripts. Real outputs included.

Convert between PNG, JPEG, WebP, AVIF, HEIC, GIF, ICO and PDF with Pillow. Measured file sizes, keeping transparency and EXIF, animated GIF to WebP, and a batch converter script.

PIL Image to NumPy array and back: shapes and dtypes for every mode, why arrays are read-only, float and int64 pitfalls, black-image bugs, and applying a matplotlib colormap. Tested code and real output.

Read an image's width and height, format, mode, channels, DPI, file size and metadata with Pillow, without decoding the pixels. Includes the EXIF-rotated size gotcha and a fast folder scanner. Tested output.

Load images into Pillow straight from a URL, raw bytes, a base64 string, a data URI or an upload, without temporary files. Handles errors, large files and streaming. Tested examples with real output.

Turn a Pillow image into PNG/JPEG bytes, a base64 string or a data URI, and send it from Flask, FastAPI or Django without saving a file. Explains tobytes() vs save(). Tested examples with real output.

Shrink JPEG, PNG and WebP files with Pillow: quality, optimize, progressive, subsampling, palette PNGs, metadata stripping and resizing, each measured on a real photo. Plus a target-size compressor.

How to display a PIL image: Image.show() and why it sometimes does nothing, inline in Jupyter and Colab, with matplotlib (and the "can not convert to float" error), and in a Tkinter window.

Convert PIL images to OpenCV arrays and back without the blue-skin bug: BGR vs RGB, alpha channels, greyscale, cv2.imread vs Image.open differences, and speed. Tested with OpenCV and Pillow, real output.

Read EXIF metadata with Pillow's getexif(): tag names, the date a photo was taken, camera and lens, GPS coordinates in decimal degrees, orientation, plus editing and stripping EXIF. Tested output.

Merge several images into one with Pillow: side by side, stacked, or as a grid / contact sheet, with images of different sizes, gaps, background colours and labels. Tested code with the real output images.

How Image.crop() boxes work (left, upper, right, lower, not x, y, width, height), centre and aspect-ratio crops, circular avatars, and auto-cropping white borders, including the JPEG noise that breaks getbbox().

Read a pixel's RGB value, get all pixels, and change pixel colours with Pillow: getpixel, load(), get_flattened_data (getdata is deprecated), NumPy, and how fast each one is on a real photo.

Paste a transparent PNG (logo, watermark, sticker) onto a photo with Pillow without black corners. When to use paste with a mask vs alpha_composite, opacity, positioning, and the "bad transparency mask" error.

Turn white (or any colour) pixels transparent with Pillow: why exact matching fails on JPEGs, tolerance and soft edges with NumPy, flood fill to keep white parts of the subject, and replacing one colour.

Cut an image into an N x M grid, fixed-size tiles, overlapping patches for machine learning, or a seamless Instagram carousel, handling sizes that don't divide evenly. Tested code with real output.

Install Pillow correctly with pip in a virtual environment, conda, Jupyter or Docker; fix "externally-managed-environment", pip install PIL failing, and old Python versions. Tested commands.

Greyscale with convert("L"), pure black and white with convert("1"), why the result is speckled (dithering), choosing a threshold (fixed or automatic Otsu), and keeping transparency. Real output.

What each Pillow image mode stores, how many channels it has, what getpixel returns, which mode each file format opens as, the difference between P and L, and conversions that silently lose data.

Read every image in a directory into a list, resize or convert a whole folder, sort file names naturally, skip bad files, avoid "Too many open files", and use all CPU cores. Tested on Linux, Windows and macOS.

Make an animated GIF or WebP from frames, read frame count and durations, extract frames, resize without losing the animation, and fix trails in transparent GIFs (disposal). Tested code, real output.

Save one or many images as a PDF with Pillow: why the pages come out huge, setting the page size with resolution, fitting images on A4 with margins, file size and quality, and transparent PNGs.

Draw shapes with ImageDraw: how box coordinates and outline width work, semi-transparent fills (and why alpha is ignored on RGB images), smooth anti-aliased circles, thick lines, polygons and labelled boxes.

Image.rotate() turns counter-clockwise and crops by default. How to rotate clockwise, keep the whole image with expand=True, fill the corners, avoid jagged edges, flip/mirror, and fix sideways phone photos.

Open multi-page TIFFs, save stacks, choose compression, keep 16-bit and float data intact through NumPy, convert TIFF to JPEG, and the 48-bit colour TIFFs Pillow silently reduces to 8 bits. Tested, real output.

Make a negative with ImageOps.invert, and fix "not supported for mode RGBA" and "mode P support coming soon": invert transparent PNGs without losing the alpha, palette, CMYK and 16-bit images. Tested per mode.

Fix "No module named 'PIL'": install Pillow (not PIL) into the same interpreter that runs your code. Exact commands for venv, Windows, macOS, Linux, conda, Jupyter, VS Code and Docker.

Image.ANTIALIAS was removed in Pillow 10. Replace it with Image.LANCZOS (same filter, works on every version since 2.7), patch old libraries, and see why pinning Pillow<10 fails on Python 3.12+.

ImageDraw.textsize() and FreeTypeFont.getsize() were removed in Pillow 10. Replace them with textbbox() or textlength(), with a drop-in helper, a centring fix and real measurements.

Why Pillow says "cannot identify image file" and how to fix it: HTML saved as .jpg, empty downloads, BytesIO not rewound, HEIC without pillow-heif, SVG. Includes a byte-sniffing diagnostic script.

ImageFont.truetype() raises "cannot open resource" when FreeType cannot find the font file. Where Pillow searches on Windows, macOS and Linux, when the name is case-sensitive, and a robust loader with fallback.

Image.fromarray raises "Cannot handle this data type: (1, 1, 3), <f4" or "<i8" when the NumPy dtype or shape has no Pillow mode. How to read the message, and the exact conversion for floats, int64 and channel-first arrays.

"cannot write mode X as JPEG/PNG" means the format cannot store that pixel mode. A tested table of which modes each format accepts, and the right conversion for F, P, LA, I, I;16, CMYK and RGBA images.

Pillow raises "image file is truncated (N bytes not processed)" when a file ends early. Why open() works but load() fails, how to detect bad files, and when LOAD_TRUNCATED_IMAGES is safe. Tested on real images.

Pillow warns above 89,478,485 pixels and raises DecompressionBombError above twice that. What the limit protects against, how to raise it safely for trusted images, and how to process huge images without it.

matplotlib's imshow() raises "Image data of dtype object cannot be converted to float" (older versions: "Image data can not convert to float") when it gets None, a path or a list instead of an image. Causes and fixes.

Why changing one channel of a Pillow pixel, or img.size, raises "'tuple' object does not support item assignment", the fixes, and the NumPy version of the same mistake: "assignment destination is read-only".

scipy.misc.imread, imresize, imsave and toimage were removed from SciPy. Drop-in replacements with Pillow, imageio and matplotlib, and the dtype, size-order and rescaling differences that silently change results.

Fix AttributeError: module 'PIL' has no attribute 'Image', No module named 'Image', type object 'Image' has no attribute 'open' and NameError: name 'Image' is not defined. Each reproduced, with the import that works.

Pillow's compiled core failed to load: cannot import name '_imaging', DLL load failed, "built for another version of Pillow", missing .so libraries in Lambda/Docker, _imagingft and "decoder jpeg not available".

Image.Resampling, Image.Transpose and the other enums were added in Pillow 9.1. Reproduced on Pillow 9.0.1, with the upgrade, a constant that works on every version, and how to find what pins an old Pillow.

Django's fields.E210 error, reproduced: when Pillow really is missing, when it is installed in a different environment, and when it is installed but broken (Django hides the real ImportError). Tested with Django 6.