cineColorQt — segments of the film on the left, the corrected frame in the middle, groups and their settings on the right.
Super 8 material shows all kinds of colour casts. They can come from the scene, from the film stock and its ageing, or from the scanner. Correcting them with the primary colour wheels of DaVinci Resolve was too laborious for me: too many controls, each affecting the others.
So the idea was to automate it. The result is a program that first analyses a film and then estimates values for lift, gamma and gain, similar to DaVinci, and goes a bit beyond that. In some cases the transfer curves that lift, gamma and gain can produce are not enough. cineColor therefore works in two stages: Stage 1 searches for the best lift, gamma and gain, and Stage 2 then adjusts each of 13 brightness bands individually to smooth out the remaining deviations.
The core idea of the automatic estimate is to find gray areas in the frames and to correct the colours so that these areas come out gray. This does not always work, and for those cases there are manual controls at several points of the pipeline. In the extreme case the user marks the areas that should be gray, as "gray picks".
The colours can also be shifted by hand, both at the input and at the output of the estimator. Shifting the output allows, for example, to make a scene a bit sunnier again than it came out of the correction.
cineColor is one of the cineTools, next to cineFlow and cineAlign.
Work starts by dropping a folder with the frames of a film onto the window. The program analyses the frames and tries to divide the film into segments of coherent colour.
All segments are first put into one large group, "All".1 In a first check, take segments whose ΔR/G or ΔB/G is marked orange out of the estimate by removing their checkmark; such a segment is then passive. Typically these are scenes of a sunset, for example. Taking these outliers out usually stabilises the estimate.
This is the moment to look at the whole film, or at least at the important scenes. The computed correction may well be usable already. cineColor then offers two ways to get the corrected material. Export LUTs … writes the computed LUTs in a format that the sister program cineAlign reads and applies. Alternatively, Export Material … renders the corrected film directly from cineColor, for instance as a folder of 16-bit TIFF files.
cineColor uses a number of keyboard shortcuts; the key h shows an overview.
If the simple approach with the group "All" does not lead to success, there are several other options. Sometimes it helps to put a scene or a sequence whose colours are off into a group of its own: right click in the segment list and choose New group for n segment(s) …, or use the button New group … above the list. A typical case that calls for this is a part of the film with both night and daylight shots. Here it can make sense to introduce a group "daylight" and a group "night time" and to have them estimated separately.
The same context menu assigns segments to a group as active or passive. The transformation is estimated only from the segments of a group that are active; the passive ones are corrected with it, but do not contribute to it. This way a single scene with many gray tones can be used for the estimate (set to active), and the resulting transformation applied to 20 or 30 other scenes that ended up on the same reel at the same time (same group, but passive).
For extreme cases in which the automatic estimate fails, cineColor offers the "Picks" mode. Here the user sets test points by hand on parts of the image that are known to be gray. If even this does not help, there is a completely manual mode, in which the colours are shifted with the mouse in the vectorscope (see below).
Each group has its own estimator, chosen with the buttons in section 1 of the group panel:
| Original | no estimate; only the manual shifts (trims) apply |
| Stage 1 | lift, gamma and gain per channel, through three gray anchors in the shadows, mid tones and highlights |
| Stage 2 | Stage 2 alone, on the uncorrected material |
| Stage 1+2 | Stage 1, then Stage 2 on its result — the default |
| Picks | from the gray picks of the group |
| LUT | an imported LUT, as it is |
Stage 1 is iterated: each pass picks the least saturated pixels anew under the correction found so far (default 5 passes). With a strong cast this helps; with a dominant scene colour such as sand or grass, the picks drift towards that colour.
Stage 2 looks for the gray cluster in each of the 13 brightness bands. A cluster that lies farther from neutral than the tolerance (default 0.06 decades) counts as coloured, and its band is not used. With too few usable bands, Stage 2 is dropped. The log below the group panel shows how many bands were accepted.
Strength scales the estimate, from 0 to 150 %. At 0 % only Trim out remains. This is probably the easiest option for fine tuning an automatich estimate - setting this value to 60% or so will reduce for example overcompensation which might occur for very sunny scenes.
The two trims shift red and blue against green, in decades, by up to ±0.15. The shift is constant up to the highlights and fades out towards white.
- Trim in acts before the estimator: the estimator works on the shifted material. It helps where a strong cast leaves Stage 2 too few gray bands. Often the estimate absorbs it; sometimes it stays in the result. The strength scales it together with the estimate.
- Trim out acts after estimate and strength. It is the last touch by hand and is not scaled by the strength.
Values other than zero are shown in orange. The checkbox active next to Trim out switches it off in the display only, to compare; the LUTs and the exported material keep it. Switching off Trim in removes it from the computation altogether. Reset sets a trim back to zero; a double click on the strength slider or its label resets the strength.
Try Trims if the Strenghtslider does not improve the result.
The vectorscope at the bottom right shows the colours of the visible part of the frame: In, the material after Trim in, and Out, the result. A click on "In" or "Out" shows or hides that cloud. The mouse wheel changes the gain, and the circles mark the colour under the mouse pointer in the image.
With only one of the two clouds shown, dragging inside the circle shifts it:
- dragging Out sets Trim out, and the image follows at once;
- dragging In sets Trim in; the estimate is recomputed when the mouse is released.
Esc while dragging cancels the drag. The distance moved does not depend on the gain of the display; one pixel is about 0.0025 decades. With both clouds shown, nothing is dragged.
The estimator Original together with dragging Out is the completely manual mode: no estimate, only the shift by hand. With any other estimator, dragging Out shifts the result of the estimate.
They are the last semi-automatic color correction option if the other algoritms fail. In this mode, you choose image areas which you know to be "gray", that is, without color.
A Shift+click on a gray area of the frame sets a gray pick; the square shows the area that is sampled. A new pick selects the estimator "Picks" for its group. With one pick the correction is a gain, with two lift and gain, from three on gamma as well; Stage 2 is not used.
Only picks on active segments count. Picks on passive segments are drawn pale and have no effect.
A right click on a pick deletes it from the current version. P and Shift+P jump to the next and previous frame with picks; with the estimator "Picks", the group panel lists these frames as well. Choosing another estimator keeps the picks, so the result with and without them can be compared.
Picks belong to a version of the group (next section). If a segment moves to another group, its picks stay with the version they were set in, and count again when the segment comes back.
A group can hold several versions of its settings. Ctrl+Y creates a new version as a copy of the current one, Ctrl+B and Ctrl+N step to the previous and next version, Ctrl+X deletes the current one. The version shown is the one that counts, also for the LUTs and the exported material.
A version holds the estimator, the strength, both trims, the Stage 1 passes, the Stage 2 tolerance and the gray picks. Segments, checkmarks and an imported LUT are the same for all versions of a group.
Different sets of gray picks are compared this way: Ctrl+Y, delete or add picks in the new version, then step back and forth.
Another option available with versions: you can have one version using Stage 1+2, and another using a LUT imported from somewhere else.
A .cube written by cineColor, from this film or another one, can be dropped onto the window:
- dropped on the image, it becomes a new, empty group with the LUT as its correction; segments are assigned to it afterwards;
- dropped on a row of the group list, it goes into that group and becomes its estimator "LUT". Segments, versions and picks stay.
When cineColor writes new LUTs, it renames the existing ones to .cube.alt. These can be dropped as well.
Right click a group and choose Copy to new group (frozen) … to keep the current state of a group as a fixed LUT before working on: an empty group with that LUT. If the earlier state turns out better, move the segments into the copy.
The analysis looks at a fixed part of each frame, so that sprocket holes, frame lines and the borders of an overscan stay out of it. Key 5 shows this area as a yellow frame on top of the current view: drag its edges to resize it, or drag inside it to move it. Pressing 5 again hides it; if the area was changed, cineColor asks whether to analyse the film again. Segments, groups and picks are kept, only the estimates are computed anew.
The area is stored with the material. New material starts with the area used last.
cineColor is a set of Python scripts; put them in one folder. It needs Python 3.11 or newer and these packages:
pip install numpy scipy opencv-python PyQt5 matplotlib
Then start it:
python cineColorQt.py
The key h lists keys and mouse actions. Most controls explain themselves in their tooltips. Beyond that, you are on your own.
This is version 4.0: working software that one person uses on his own films, published in the hope that it is useful to others.
No support is promised. Bug reports are welcome and will be read.
GPL-3.0-or-later. See LICENSE.
Commercial licences are available for use cases the GPL does not cover. Enquiries: license@pixelcircus.com
Copyright (C) 2026 Dr. R. Henkel
Footnotes
-
Segments that the analysis finds unsteady (roll ends, splices, flashes; marked
~after their number) start in "1:1" instead, without correction. ↩
