AI color grading is software that analyses your footage and proposes color changes automatically, instead of waiting for you to move wheels and curves. In practice it covers three different jobs: correcting exposure and white balance, transferring the color of a reference image onto a shot, and matching shots to each other so a scene cuts together.
Those three jobs have different maturity levels, different failure modes, and different names in every tool's marketing. This guide separates them, shows where the maths actually comes from, and says which parts still need your eye.
What AI color grading actually does
1. Automatic correction and balance
The oldest and least glamorous mechanism. The software reads the image statistics (how the luma sits across shadows, midtones and highlights, where the neutral point falls, how saturated the frame is) and pushes exposure, white balance and contrast toward a technically sane result. A chip chart makes this exact rather than inferred: DaVinci Resolve's own documentation describes it as "click on the color match icon in the toolbar, select the type of chart used, line it up on the image and click match!" (Blackmagic Design, read 12 September 2026).
Automatic correction aims at neutral, not at a look. That is why footage can come out of an auto pass technically defensible and creatively flat. It is also why it is a good first step on log footage, where the real problem is usually that nothing has been normalised yet. If you are unsure what your camera's log curve is doing to code values before any of this runs, our log format reference computes it for nine formats.
2. Reference and look matching (color transfer)
This is the mechanism people mean when they say a tool is "AI". You give it a picture (a still from a film, a graded frame of your own, a photograph a client sent) and it moves your shot's color toward that picture.
The idea is older than the current wave of tools. Reinhard, Ashikhmin, Gooch and Shirley published "Color transfer between images" in IEEE Computer Graphics and Applications 21(5), pages 34-41, September 2001; their abstract describes using "a simple statistical analysis to impose one image's color characteristics on another" (University of Bristol research record). The method works in a color space where the axes are close to uncorrelated, so the mean and spread of each channel can be re-shaped separately.
Pitié, Kokaram and Dahyot generalised that in "Automated color grading using color distribution transfer", Computer Vision and Image Understanding 107, pages 123-137, 2007 (DOI 10.1016/j.cviu.2006.11.011, Maynooth University record). Instead of matching a mean and a standard deviation, they transfer the whole color distribution with an iterative, parameter-free algorithm, then post-process to keep the original image's gradient structure and avoid raising noise.
Modern tools layer learned models, face and sky awareness, and gamut handling on top, but the shape of the problem has not changed: a reference is a distribution, your shot is a distribution, and the tool finds a mapping between them. Knowing that tells you when it will work. A reference that shares your shot's content and lighting transfers cleanly. A reference dominated by a color your shot does not contain transfers badly, because the mapping has to invent a correspondence that is not there. We wrote up the specific case of pulling a look off an edited photograph in our guide on how to match video to a still.
3. Shot-to-shot matching
Matching is color transfer where the reference is another shot of yours rather than an outside image. The goal is agreement: the same exposure, the same white point, skin in the same place, across every cut in a scene.
Shot matching suits automation because much of the first pass is measurement, while the editor still decides what the scene should look like. A match can underperform when the source is badly exposed, the input transforms differ, or the reference clip was never corrected. If you have hit that wall in Resolve, see why Resolve's Shot Match underperforms. For the full preparation, matching, and motion checks, use the AI shot-matching workflow.
What it cannot do yet
Limits, because every one of these will cost you time if you plan around the marketing instead.
Creative intent. No current tool knows that the third act should cool off, or that this particular face should sit half a stop brighter than the meter wants. Color transfer moves a shot toward a reference; choosing the reference is the creative act, and it stays yours.
Mixed light. A frame lit by a window and a tungsten practical has two neutrals in it. Distribution matching sees one image and produces one answer, so the compromise it picks is often not the one you would pick. Local work, a mask, or a targeted correction on that shot is still the fix.
Skin. Skin is where human tolerance is narrowest and where automatic answers most often need adjustment, especially under mixed light or heavy makeup. Treat any automatic result on faces as a starting point to be checked, not a finished decision.
Extreme color. A dominant neon sign, a red emergency light, a colored gel across most of the frame: these distort the statistics the match is computed from, and the tool may try to "correct" the very thing you wanted.
Review across the edit. Grade one shot, then use Match shots to match the remaining shots to it. Review faces, exposure and adjoining cuts, refine the result with the creative controls, and export the finished sequence.
The four operations in Leumos
| Operation | Starting point and result | Reference-match allowance |
|---|---|---|
| Reference match | A still image guides the selected shot's color and contrast; inspect it in motion and refine the grade. | A submitted reference match uses the monthly allowance. |
| Copy grade | Transfer grading settings from a graded shot to other shots; check the receiving shots' exposure and color. | Copying settings itself uses no reference match. A separate reference-match run is metered. |
| Match shots | Use a graded anchor shot to match the remaining shots, then review faces, exposure and cuts. | Some shots can be handled without a reference-match run; shots sent for reference matching use the monthly allowance. Review the run's usage. |
| AI Color Correction | Analyze a shot, or run correction across the timeline, to adjust its technical balance without a reference still. Review and refine the result. | The correction operation itself uses no reference match. |
Import supported video or a Final Cut Pro 7 XML edit from Premiere or Resolve with its source media. See the online grading workflow for the path through review and finished-video export.
The tools filmmakers actually compare
These are the products that appear when people search for AI color grading and video work, plus the two editors most filmmakers already own.
| Tool | Where it runs | How you give it a look | Works on video | Price entry point |
|---|---|---|---|---|
| Colourlab AI | Desktop app (AI 4, Apple Silicon Macs); AI 3 also as a Resolve plugin | Automatic engines, reference stills, and prebuilt looks | Yes | AI 4 prices on its pricing page, read 21 October 2026 |
| fylm.ai | Browser and cloud | Look design and film-emulation tools; its own page describes "neural-network based, smart film emulations" | See vendor page | fylm.ai |
| DaVinci Resolve | Desktop application, macOS, Windows, Linux | Chart-based Color Match, Shot Match against another clip, plus hands-on grading | Yes | Free version, paid Studio version: blackmagicdesign.com |
| Adobe Premiere Pro, Lumetri | Desktop application inside Premiere Pro | LUTs, presets and Lumetri's own color controls on the clip | Yes | Subscription: adobe.com |
| Imagen Video | Cloud service for editors and videographers | Automatic correction with per-user profiles; its own page says it "matches every camera, fixes every skin tone" | Yes | imagen-ai.com/video |
| Leumos | Browser, no desktop install | Selected-shot Reference match, Match shots from a graded shot, Smart Grade, AI Color Correction and creative controls | Yes | Free plan, then Lite and Studio: leumos.ai/#pricing |
Checked 12 September 2026, and the Colourlab price on 21 October 2026. Prices change, so each row links the vendor's own page rather than freezing a number here. The fylm.ai, Adobe and Imagen sites could not be checked that day, so those rows link out and quote only wording those vendors publish about themselves; the Colourlab and Blackmagic cells were read directly from the pages linked. Longer write-ups of two of them: Colourlab AI alternatives and fylm.ai review for video. A wider ranking sits in the best AI color grading tools.
How reference matching works in Leumos
Leumos is a browser color-grading studio with a timeline and finished-video export. Reference matching runs on one selected shot at a time. You can refine the shots, review the sequence and render the completed grade.
- Import and detect cuts. Leumos detects cuts so you work with individual shots in the Studio timeline.
- Choose one shot and one still. You select a shot and a still-image reference, then run the match.
- Review it in motion. The match is reviewed as moving footage, not as a frozen frame, because a match that holds on one frame can drift across a move.
- Refine. Smart Grade offers a shortlist of looks with a strength control; AI Color Correction handles balance; the Skin tone control selects skin automatically, lets you inspect and refine the selection mask, then adjusts hue, warmth, color boost, brightness and contrast. Built-in waveform and vectorscope displays let you check exposure, color balance and saturation rather than trusting a monitor you have not calibrated.
- Carry it across. You can copy a grade to another shot as a starting point, then check exposure, white balance and the result in motion. Reference matching runs on one selected shot at a time.
- Export. Graded shots export as a timeline. Resolution, bit depth and ProRes 422 HQ availability depend on the plan.
Leumos exports finished rendered graded video, rather than grade data, AAF or CDL. Its Edit handoff export gives each graded shot as its own file with handles, plus a Final Cut Pro 7 XML to relink them in Resolve or Premiere. Paid plans can download the grade as a LUT. If the production requires a camera-original conform or editable grade handoff, use an appropriate editor for that distinct step. BRAW and ProRes RAW also need debayering in a compatible editor before import; Leumos does not decode camera RAW.
The Free plan lets you try the workflow on your own footage: 10 reference matches a month, 10 projects, 1080p 8-bit export without a watermark, and 1 GB of cloud footage storage shared across the account. Local originals stay on your device. Scene detection and reference matching send material for server processing; export uploads original footage for rendering. See the online video-grading workflow, check how footage is handled, or open the Studio. Current allowances are in the pricing section.
A working order for a mixed-camera edit
Most disappointing results come from running the clever step before the boring one. The order below costs nothing and removes the majority of failures.
Normalise before you match. Tag each clip's input transform so every camera lands in the same working space. A match computed across two clips that are still speaking different languages is a match to an artefact. If your source is log and you are unsure what its curve does, check it against the manufacturer's published values rather than a forum LUT.
Correct before you grade. Fix exposure and white balance on the shot you intend to use as your reference. Whatever is wrong with that shot becomes the target every other shot is pulled toward, which is the single most common reason an automatic pass "does not work".
Pick a hero shot on purpose. The best reference is the shot with the most typical content and lighting for the scene, not the prettiest one. A close-up of a face against a window is a bad target for a wide of an empty room.
Match, then look at motion. Judge the result on moving footage. A match that holds on the frame you paused at can drift through a pan, an iris change, or a subject walking from sun into shade.
Check on scopes, not on vibes. A waveform tells you whether two shots actually share a black point; a vectorscope tells you whether skin sits on the same line. Your eyes adapt within seconds of looking at a graded frame, and scopes do not.
Leave the hard shots for last. Mixed light, heavy neon, and faces under colored practicals are where your time is worth most. Let automation clear the ordinary shots so you still have attention left for the three that need it.
FAQ
Is AI color grading any good?
For correction and matching, yes: those are measurement problems and software measures well. For creative look design, it is a fast way to get a starting point from a reference you choose, not a replacement for the choice. Expect a strong first pass and a human review, particularly on faces and mixed light.
What is AI shot matching?
Shot matching is making every clip in a scene agree on exposure, white point and skin, so cuts do not draw attention to themselves. AI shot matching automates the measurement: the tool compares one shot's color distribution to a reference clip's and computes the correction that lines them up.
Can AI match two different cameras?
Often, yes, once both are normalised from their own log or gamma into the same working space. What defeats it is not the sensor difference but the setup: wrong input transform, footage exposed several stops apart, or one of the two clips being graded before it was corrected. Normalise first, then match.
Does AI color grading work on log footage?
Yes, and log is usually the better input because it retains more range to work with. Tag the correct input transform first, since a match computed on wrongly interpreted log is a match to the wrong picture. Our log format reference computes code values and stop ladders for nine formats.
AI color grading versus LUTs: what is the difference?
A LUT is a fixed mapping applied identically to every clip you put under it. A reference match is computed per shot from that shot's own content, so two clips shot minutes apart get different corrections to reach the same result. LUTs are for a repeatable look; matching is for agreement between shots.
Will AI replace colorists?
Not on the evidence of what current tools do. They compress the correction and matching passes, which is where the hours went, and leave the decisions about what a scene should feel like untouched. The work that shifts is the mechanical part.
How do I start without buying anything?
DaVinci Resolve's free version covers chart-based Color Match and Shot Match on the desktop. In the browser, Leumos's Free plan includes ten reference matches a month with 1080p export. Both are enough to decide whether automatic matching helps your particular footage.
Related reading
- AI shot matching for video: prepare the sources, choose a reference, and inspect the result in motion
- Match video to an edited still: when the reference is a retouched photograph
- Why Resolve's Shot Match underperforms: the preparation that fixes most failures
- The best AI color grading tools: the wider field
- Colourlab AI alternatives and fylm.ai review for video
- Introducing Leumos: the story behind this workflow
- Color correction vs color grading: the two passes AI tools automate differently
- How to color grade video: the seven-step order that makes matching work
- Free color grading software: what is actually free, tool by tool
- Log footage explained: what log is and how to grade it
- How to read a vectorscope: hue, saturation and the skin line
- How to read a waveform monitor: exposure and matching on the scope
- Cinematic color grading: how film looks are built
- Color grading examples: a measured before and after
- Teal and orange color grading: the complementary look done well