QuickWB

Open, drag or paste an image in, white-balance it, and export it or copy it to the clipboard. A small single-window desktop app (PySide6).

Works on one image or on a batch, and the selection decides which: Separately gives each image its own correction, Together computes one shared correction for the whole selection so a set stays consistent.

What it does

• Drop, open, or paste images (PNG, JPG, TIFF, BMP, WebP). Drag several in at once, or paste with Ctrl+V — a screenshot, a copied image, or copied files. A drop on a slot fills that one; a drop anywhere else in the gallery opens as many new slots as it needs. Opened files are named after the file; pasted ones become clipboard #1, #2, … so two screenshots are never the same label.
• One image at a time, or a gallery. Single shows the current slot at window size, with ◀ / ▶ (or the arrow keys) and a 3 / 7 counter naming the file. Grid lays every slot out side by side and scrolls, and there the arrow keys walk it the way it looks — left and right along a row, up and down between rows; as soon as a second image is on screen the app switches to it for you. Double-click any image for a full-screen view — scroll to zoom, drag to pan, ◀ / ▶ for the next image, Esc to close. Nothing is drawn on top of the picture, so it screenshots clean: stepping to another image, or closing and re-opening, is what brings the zoom back to a plain fit. There is no counter up there, so the same 3 / 7 line flashes over the image as you land on it, then fades.
• Slots. + add another slot appends an empty drop zone and takes you to it; Delete removes any selected slot, and the last image clears back to a drop zone instead of leaving you stuck with it. Click to select, Ctrl+click to add, Shift+click for a range, Ctrl+A for the lot, and a click on the background (or Deselect) to select nothing. The slot you are on is ringed — solid when it is part of the selection, dashed when it is not — and the arrows walk from there; every control acts on the selection.
• Ctrl+Z undoes the last thing that changed the images: an open, a paste, a delete, a white balance, a slider — and says which, so you can see what came back. Ctrl+Shift+Z (or Ctrl+Y) puts it back: images you undo away are kept, so deleting and changing your mind costs nothing. Selecting, switching view and Show original change nothing, so they leave the history alone.
• White point holds everything that decides the neutral colour. Loading an image corrects nothing: you keep seeing it as it came in until you press Auto (estimate the neutral colour from the image itself) or Pick… (an eyedropper — click, or drag a rectangle over, something that should be grey or white on any selected image; a rectangle averages everything inside it). Those two are also what re-runs a balance, so there is no separate apply step, and everything below them stays greyed out until one has run. The scope above them is a readout, not a control: it follows what you selected — one image is balanced on its own, several get one shared white point, and a pick then covers all of them. The De-cast slider sets how hard Auto looks (brightest pixels only ↔ whole frame) and re-runs it as you drag. Underneath, the colour the last pass settled on: a swatch, its hex value, and where it came from — auto or picked on its own, shared auto (1, 3-5) or shared picked (all) when several images were balanced together, naming the images it was computed from — that is history, so re-balancing one of them on its own does not change what the others say. Slots are counted down the grid and renumbered as you move things about; an image that has since been deleted or replaced shows as ex2, the slot it came from. varies when the selection does not agree, and white balance not computed yet before the first pass.
• Fine-tune with temperature, tint, brightness and contrast; Reset puts the four of them back.
• Show original is an on/off switch that previews the selected images untouched — select one and it compares just that one, next to its corrected neighbours. It waits until there is a correction to compare against, and exports match what you see: an image you never balanced is saved as it came in.
• The Image panel names the ringed image and gives its size in pixels, or counts the selection when several are picked.
• Save (PNG/JPEG/TIFF; a folder when several) and Copy — Ctrl+C copies the selection: one image as a bitmap, or several as files you can paste into PowerPoint as separate pictures.

Every slider has an editable value box (Lightroom-style scales: de-cast 0–100, temperature / tint / brightness / contrast −100…+100) — drag for whole numbers, or type a fraction like −2.5 the slider cannot land on.

Live edits run on a downscaled preview so sliders stay snappy; export and clipboard copies always re-render at full resolution.

Install (Windows)

No admin rights, no PATH or registry changes, and nothing written into your user profile.

1. On the GitHub page: Code ▸ Download ZIP.

2. Right-click the ZIP ▸ Properties ▸ tick Unblock ▸ OK. Windows flags everything that came from the internet, and without this step SmartScreen stops the launcher with "Windows protected your PC".

3. Extract it somewhere it can stay — C:\Users\Public\QuickWB works for every account on the machine; your Documents folder is fine too.

4. Double-click QuickWB.bat. Nothing is asked: the Python runtime goes into the shared folder C:\ProgramData\PyApps, where every tool set up this way re-uses the same uv.exe, Python and package cache — files common to two of them are stored once (hard-linked), so a second tool only costs the libraries it does not already share.

   Prefer it self-contained? Run QuickWB.bat private from a terminal (or 2) the first time, and everything lands in C:\Users\Public\QuickWB, a single folder that uninstall.bat deletes whole.

   Either way it downloads uv, a matching Python and the dependencies — about 260 MB on disk, a minute or two on a normal connection — and puts a QuickWB shortcut on your Desktop.

5. From now on use the Desktop shortcut. Leave the extracted folder where it is; the shortcut points at it.

The choice is made once: from then on the folder on disk is the memory, so the Desktop shortcut and a plain double-click find it and start straight away — no argument needed again. Set QUICKWB_RUNTIME to force some other location.

uninstall.bat reverses all of it: the whole folder in the private case, only QuickWB's own virtual environment in the shared case, so anything else using that folder keeps working.

Run it from source (any OS)

git clone https://github.com/jakubtoczek/quickwb
cd quickwb
uv run --project . quickwb
# or, with your own environment:
pip install -e .
python -m quickwb

python misc/smoke_test.py   # gallery / white-balance self-check
bash   misc/root_test.sh    # where the launcher decides to install

Requires Python ≥ 3.11.

How the white balance works

Two steps, both textbook colour-constancy methods.

1. Estimate the white point — the colour in the frame that should be neutral. QuickWB blends two classic estimators; the De-cast slider is the mix:

De-cast	Estimator	Assumption	Behaviour
0	White-Patch (Max-RGB, from Land's Retinex)	the brightest pixels are a white surface	subtle — highlights are often near-neutral already, so it under-corrects
100	Gray-World (Buchsbaum)	the whole scene averages to grey	strong — removes a heavy cast, but greys out a scene that is genuinely one colour

The default 50 suits most photos. Both estimators use a percentile rather than a hard max or mean, and drop near-black pixels, so one blown highlight or a dark border can't drag the estimate. They are the two ends of one family — the Shades of Gray Minkowski-norm framework, where p=∞ is White-Patch and p=1 is Gray-World — so the slider is just moving along that axis.

Pick… replaces the estimate with the mean colour of what you point at — a small patch under a click, or everything inside a rectangle you drag. The manual version of the same thing, and the most reliable option when the shot contains a grey card, a white wall, or a sheet of paper.

2. Apply a von Kries diagonal transform — multiply each channel by target / white_point, so the estimated white point lands on near-white (245, 245, 245). One gain per channel, no matrix: hues shift only as much as removing the cast requires. This is what GIMP's white-point eyedropper does.

Temperature and tint nudge the white point (red↔blue, green↔magenta) before the gain is computed, so they behave like a camera's white-balance dial rather than a colour wash over the top. Brightness and contrast are a plain tone curve applied afterwards.

In Together mode the estimators run once over the pooled pixels of the whole selection (subsampled past 400k pixels), and every image gets that single white point.

References:

• E. H. Land, The Retinex Theory of Color Vision, Scientific American 237(6), 1977 — White-Patch / Max-RGB.
• G. Buchsbaum, A spatial processor model for object colour perception, J. Franklin Inst. 310(1), 1980 — Gray-World.
• G. D. Finlayson & E. Trezzi, Shades of Gray and Colour Constancy, Color Imaging Conference, 2004 — the family both belong to.
• J. von Kries, Chromatic adaptation (1902) — the per-channel diagonal gain.

The maths lives in src/quickwb/wb.py; run it directly for a self-check:

python src/quickwb/wb.py