Color correction is the technical pass: you fix exposure, white balance, contrast and black level so a shot reads as an accurate record of what was in front of the lens, and so every shot in a scene agrees with every other. Color grading is the creative pass: you push colour and contrast away from accurate on purpose, to serve a mood, a period, a genre or a brand.
Adobe draws the line the same way on its own Premiere Pro colour page: "Color correction consists of correcting issues and balancing shots, so they look natural and unified. Color grading consists of adjusting color and contrast to serve a creative or stylistic choice" (adobe.com, read 12 September 2026).
That is the whole distinction. Everything below is about why it keeps getting blurred, and what changes in your workflow once you hold the two apart.
Why the terms collapse into each other
Three reasons, and none of them is that people are sloppy.
The same controls do both jobs. A lift wheel nudged a quarter-stop to kill a blue cast is correction. The same wheel dragged hard into teal is grading. The tool cannot tell you which one you are doing; only your intent can.
The job titles merged. On a feature, correction and grading are separable stages with separable people. On a two-person corporate shoot the editor does both in one sitting, calls the whole thing "the grade", and is not wrong in any way that matters to the client.
Clients only have one word. "Can you colour it?" covers everything from "the dress has gone yellow" to "make it look like Dune".
Holding the terms apart is still worth it, because the two passes fail differently and need different evidence that they worked.
Where each one sits in a real pipeline
The order below is not a style preference. Every step depends on the one before it being finished, and most disappointing grades are a step done out of sequence.
1. Normalise. Tell the software what the footage is. Log footage carries a manufacturer's transfer curve; until you apply the matching input transform or colour space transform, nothing you do downstream is operating on the picture you think it is. This is not correction yet — it is putting every camera into the same language. If you are not sure what your camera's curve does to code values, our log format reference computes them for nine formats.
2. Correct. Now fix the shot. Exposure to where you want it, white balance neutral, black point set, contrast sane. The target here is verifiable, not aesthetic: a waveform that sits where you meant it to, skin near the vectorscope's skin line, a neutral wall that reads neutral.
3. Match. Make every shot in the scene agree with the others. Still correction — measurement, not taste — but a distinct pass with a distinct question: does this cut draw attention to itself? A scene can be full of individually correct shots that still flicker in colour temperature across the cuts.
4. Grade. Only now do you decide what the scene should feel like. Cool the night exterior, warm the kitchen, crush the blacks, desaturate the greens. You are pushing away from accurate deliberately, and you have a stable, agreed base to push from.
5. Finish. Secondaries, skin, vignettes, any shot-specific repair, then the output transform and the export.
Correction comes first for a structural reason, not out of ceremony. A look built on an uncorrected shot is welded to that shot's specific errors; apply it to the next clip and it falls apart, because that clip's errors are different. Correct first and the look becomes portable.
What correction looks like, concretely
Take an interview shot in a meeting room, window behind the subject, fluorescent tubes overhead.
Before: the face sits low on the waveform, around 35 IRE, because the camera metered for the window, which is clipped flat at the top of the trace. Skin on the vectorscope points about ten degrees off the skin line toward green, because the tubes throw a cast the auto white balance split the difference on. Blacks sit at 8 IRE instead of 0, so the image looks milky.
After correction: the face is lifted to around 55–60 IRE, the black point is brought to 0 without crushing the shadow detail that was there, and the green is pulled out until skin lands on the skin line. The window is still clipped, because correction cannot invent data the sensor never recorded. Nothing about this is a look. It is a shot that now tells the truth about the room, and one you can safely build on.
What grading looks like, concretely
Same shot, now graded for a corporate piece that wants to read as warm and confident.
The contrast curve is steepened slightly, so midtones separate and the image stops looking flat. A touch of warmth goes into the highlights and a touch of blue into the shadows. Greens in the background plant are desaturated and pushed toward yellow so they stop competing with the face. Saturation comes up a few points. The face itself is protected, because the moment skin drifts off the skin line a viewer reads it as wrong rather than as a look.
The waveform after grading no longer says "accurate". It says "this is the picture we chose". That is the point.
How the split maps to the tools you already have
| Stage | DaVinci Resolve | Premiere Pro, Lumetri | Leumos |
|---|---|---|---|
| Normalise | Input colour space transform, or an input LUT | Input LUT and Premiere's colour management | Named input transform for your source log format |
| Correct | Auto colour button, primary colour wheels, chip-chart Color Match | Basic Correction, Auto Color, Color Match | AI Color Correction, plus manual controls |
| Match shots | Shot Match to This Clip | Color Match between shots | One selected shot matched to a still reference |
| Grade | Curves, custom curves, primaries pushed for look | Creative tab, Lumetri looks, LUTs, Color Wheels | Smart Grade shortlist with a strength control |
| Secondaries | Qualifier palette, Power Windows | HSL Secondary | Skin tone control |
| Verify | Waveform, vectorscope, parade | Lumetri Scopes | Built-in waveform and vectorscope |
The Resolve cells are the vendor's own wording. Blackmagic describes the primary wheels as adjusting "balance and brightness in overlapping tonal ranges called lift, gamma and gain" and the qualifier palette as selecting "part of an image based on hue, saturation or luminance". Its correction instruction is to "click the auto color button at the left of the color wheels palette to balance the shot's color and contrast"; its matching instruction is "Select a clip, right click on another clip, and select 'Shot Match to This Clip'" (blackmagicdesign.com, read 12 September 2026).
Adobe's page splits its own panel along the same seam. Basic Correction is described as "Make basic corrections to exposure, white balance, and contrast in a snap with Auto Color. Preserve natural skin tones and automatically match color between shots with Color Match." The creative half is "Take advantage of Lumetri looks to emulate different film stock and give your footage a timeless cinematic look" (adobe.com, read 12 September 2026). One panel, two named sections, correction above grading — which is the pipeline drawn in the interface.
Leumos is ours, so treat the last column as disclosure rather than recommendation. Its published capability list names AI Color Correction and manual controls for the correction pass, reference-image matching on one selected shot at a time, a Smart Grade shortlist for look, a dedicated Skin tone control that "automatically select[s] skin, inspect[s] and refine[s] the selection mask", and built-in waveform and vectorscope displays (leumos.ai/llms.txt, read 12 September 2026). Automatic reference matching of a full timeline is not available in this release. Plans are on the pricing section.
Is a LUT correction or grading?
Both, depending on which LUT.
A conversion LUT — S-Log3 to Rec.709, C-Log3 to Rec.709 — is normalisation: a fixed mapping that undoes a known transfer curve. A translation, not a decision, and it belongs at step 1. The catch is that it bakes in a tone map a colour space transform would leave you control over, which is why the same log clip looks different through three technically correct routes: why CST, ACES and LUTs look different and S-Log3 to Rec.709 conversion.
A look LUT — a film emulation, a creative preset, a pack you bought — is grading. Somebody's fixed creative decision, applied identically to every clip you put under it.
The failure is stacking a look LUT on uncorrected footage. The LUT treats every shot the same; your shots are not the same; so it produces a different answer on each. Correct first and the LUT behaves like a look instead of a lottery.
Where automation actually helps
Correction and matching are measurement problems. There is a right answer and it can be checked on a scope, which is why the automatic features in every colour tool — auto colour, chart match, shot match, reference match — cluster on the correction side of the line rather than the grading side.
Grading is a decision problem. Software can propose a look from a reference you supply, and a good proposal saves real time, but choosing the reference is the creative act and it stays yours. The longer version of that argument, with the colour-transfer research it rests on, is in how AI color grading works.
So: automate the pass with a checkable answer, and spend what you get back on the pass that has none. Before you call anything graded, confirm each clip carries its correct input transform, that the shot you matched everything to was corrected first, and that skin still sits on the vectorscope's skin line. Judge it in motion, not on a paused frame — a match that holds on one frame can drift through a pan. If something looks wrong and you cannot name what, why your grade looks off works through the causes in the order that finds them fastest.
FAQ
Do you color correct before grading?
Yes, and the order matters more than most people expect. A look built on an uncorrected shot is welded to that shot's specific errors, so it falls apart on the next clip. Correct and match first, then grade, and your look becomes portable across the scene.
Is a LUT color correction or color grading?
Depends on the LUT. A conversion LUT that maps a log curve to Rec.709 is normalisation, done before correction. A creative or film-emulation LUT is grading — somebody else's fixed look applied identically to every clip. Stacking a look LUT on uncorrected footage is the common mistake.
Can AI do color correction?
Yes, better than it does grading. Correction and shot matching are measurement problems with checkable answers, which is why automatic features cluster there: auto colour, chart match, shot match, reference match. Creative look design still needs you to choose the reference. Check every automatic result on scopes and on faces.
Is color correction the same as color balancing?
Close, but balancing is one part of correction. Balancing means getting the neutral point right so whites read white and skin reads natural. Correction also covers exposure, black level and contrast, plus making shots agree with each other across a scene.
Which takes longer, correction or grading?
Correction usually takes more hours across a project because every clip needs it, while grading is decided once per scene and carried across. That is also why correction is the pass worth automating first — it is the bulk of the mechanical work, not the bulk of the creative work.
Do I need to correct footage that already looks fine?
Usually yes, at least to normalise it. Footage that looks fine on your laptop may still sit on a different black point or white point from the shot it cuts against, and that mismatch is invisible until the two are next to each other. A waveform tells you in two seconds.
Related reading
- How AI color grading works — the three jobs, the research they rest on, and the limits
- Why your grade looks off — diagnosing a picture you cannot name the problem with
- S-Log3 to Rec.709 conversion — LUT versus colour space transform, and what each breaks
- Why CST, ACES and LUTs look different — three correct normalisations, three pictures
- Why cameras don't match — the post-only repair when a reshoot is not an option
- Log format reference — computed code values and stop ladders for nine formats