A creative LUT will not fix mismatched shots

Dropping a "film" LUT on a mixed-model timeline feels like a decision. It is a filter. Every clip still has its own white balance, its own exposure, and its own idea of what skin is. The LUT then multiplies those private grades by the same curve. The mismatch is still there, wearing the same coat.

This is the shot-matching error that costs the most time, because it looks like progress. The cut from clip A to clip B still jumps, only both sides are teal, or both sides are print-film orange. Reviewers will say the grade is heavy. The grade is not the problem. The inputs were never on the same baseline.

The order that actually works is unfashionable: neutralise every clip, match to a chosen master shot, then grade for look. A creative LUT is step three. Used as step one, it hides the work you still have to do.

What a look LUT is for

A LUT is a baked mapping from input colour to output colour. A creative, or "look," LUT encodes a taste: contrast shape, a lift in the highlights, a push toward a print stock. It assumes the picture you feed it is already in a known state: a neutral, correctly exposed, white-balanced image in a working space the LUT was built for.

If the assumption holds, six clips through one LUT look like six clips from one camera. If it does not, they look like six different cameras with the same filter.

Generated video fails the assumption by construction. Each clip is an independent sample. The sampler picks a white point and a skin rendering as part of making the picture. There is no shared camera, no shared sensor profile, no shared CDL from set. Veo 3.1 and Kling O3 Pro will not even fail the same way. One will sit warm and smooth. The other will sit cooler, with a different red in the face. A look LUT applied to both does not average them. It maps each lie through the same function and preserves the distance between the lies.

A style preset at generation time is a cousin of this mistake. It is a named look the model applies while sampling, not a substitute for matching afterwards. Presets are also model-specific: the same label on two families is not the same grade. Locking a preset is useful on one model. It is not a LUT, and it does not discharge the match.

The mismatch the LUT cannot see

Take two talking-head clips, same prompt, different models. On a monitor, clip A looks a little gold, clip B a little magenta. You drop a cinematic LUT on the track. Both now have crushed blacks and a cool shadow. The gold clip's skin goes brick. The magenta clip's skin goes mauve. The cut is worse than before, because contrast has increased and the hue error is riding a steeper curve.

That is what "preserves the mismatch under a shared tint" means in practice. The LUT is not ignoring the error. It is amplifying whatever you fed it.

Three tells that you applied the look too early:

• Skin hue still rotates at the cut, even though both shots now "have a grade."
• Neutrals (a shirt, a wall, a product card) do not agree. One clip's grey is greenish, the other's is pink, and both are sitting under the same film curve.
• You keep making per-clip LUT opacity or mix tweaks. That is shot matching, done with the wrong tool, after the look has already been baked in.

High CFG makes the starting mismatch larger. Oversaturation and hard contrast are the visual signature of too much guidance. A look LUT on top of that is two stylisations stacked. Flatten the oversaturated clip toward a neutral baseline first.

Do not try to "counter" a warm clip by picking a cooler LUT, and a cool clip by picking a warmer one. That is four looks. You want one look, on matched pictures.

Neutralise, match, then look

The whole grade is three passes, in this order, with the look parked until the third.

Pass 1: neutralise per clip. White balance, exposure, contrast, against scopes, one clip at a time. Put a known neutral on centre. Put skin in the same neighbourhood on the vectorscope. Put the face in the same luma band on the waveform. You are stripping each model's private grade, not making the shot pretty. If a clip cannot be pulled to a boring baseline without falling apart, it is the wrong clip for this scene. Regenerate it on the master shot's model rather than inventing a salvage grade.

Pass 2: match to a master. Pick one shot whose skin and light you will live with. Usually a mid-shot, front-lit, no hero grade, no hard practical dominating the frame. Every other clip is pulled toward that master, not toward its own favourite still. This is still not a look. It is "these were the same camera." Some colour tools can derive a .cube from a reference still of the master and apply that as a match LUT. That match LUT is not creative. It is a translation into the master's photometry. Use it as a starter, then finish by eye on the cut, because a still cannot see the jump.

Pass 3: grade for look. Now the creative LUT, or a manual look, goes on a layer that sits on top of the matched sequence. One LUT. One mix value. If a clip still breaks, the break is in pass 1 or 2. Do not fork the look to hide it.

The video editor is where the sequence is assembled. preview: true gives a free 480p pass with a short per-user cooldown, which is enough to confirm the cut before you take a still sequence into a grading tool. The final export is charged once, regardless of clip count. Do the LUT work on the exported, matched clips, not on the preview file. The preview is 480p and it is not your master.

A workable stills-based match, if you do not have a dedicated matcher:

1. Export a reference frame from the master and from the clip you are pulling.
2. In a stills grade, match the clip frame to the master frame: white, then exposure, then a gentle sat match. Save that as a .cube if your tool will.
3. Apply the cube to the moving clip.
4. Watch the cut. Still grades lie about motion and about mixed lighting. Trim on the timeline until the jump is gone, or admit the clip needs a regenerate.

Building a look that survives the cut

Once the match is done, the look is allowed to be opinionated. A few constraints keep it from undoing the work.

One look per scene, not per shot. If the story changes rooms, you may change looks at the scene boundary. You may not change looks at every reverse.

Do not stack a generation-time preset, a timeline film stock, and a delivery LUT. Pick one stylisation layer after the match. The rest is how you get the burnt, over-processed "AI grade" that reviewers reject without being able to name.

Keep skin as a constraint, not a casualty. If the look pushes faces off the skin-tone line, the look is too strong for generated skin, which already sits closer to plastic than photographed skin. Back the LUT mix off until faces hold. Contrast you can add with a gentle S-curve after the LUT; hue you cannot recover if the LUT has rotated it.

Deliver from the matched, looked-at master, not from a social re-export. Platforms will re-encode. A look that only works on your monitor at full bit depth will collapse on upload if it was holding the match together with fine hue differences. If the sequence only matches with the LUT at 100% and falls apart at 70%, the match was never done.

When you generate the next batch, shrink this pass at the source: one model per scene, identical lighting language in every prompt, and a reference image for the location if the set has to hold. Colour matching is the catch, not a plan for shooting on four samplers and hoping a cube will make them one camera.

FAQ

Can I use the same creative LUT I use on real footage?

Yes, after the match, if the working space matches what the LUT expects. The LUT does not know or care that the pixels were sampled. What it cannot do is the conversion from "Veo's idea of tungsten" into "your camera's Rec.709." That conversion is pass 1 and 2. Feeding a camera LUT a raw generation is how you get the brick-and-mauve problem.

Is a match LUT the same thing as a look LUT?

No. A match LUT (including one derived from a master still) is a translation: make clip B's photometry look like clip A's. A look LUT is a style: make this already-correct picture look like a print stock. Applying a look as if it were a match is the defect this post is about. If your tool only has one LUT slot, do the match with wheels and curves, then put the look in that slot.

Why not just regenerate every shot on one model?

Do that when you can. A single-model scene is the smallest colour problem you can have. Mix models only when the jobs are genuinely different, and budget the three-pass grade as part of that choice rather than as an afterthought at delivery.

Does a higher CFG make the LUT work better?

The opposite. High guidance exaggerates the gap between the prompt-conditioned look and a natural image: more saturation, harder contrast, more plastic skin. That is a worse starting point for any LUT. If a model exposes CFG, use the middle band the model actually wants, then match, then look. Distilled and turbo variants often want much lower guidance than their full siblings; copying a number from a different card will burn the picture before a LUT ever touches it.