Read Skin Tones on the Vectorscope
Generated faces sit off the skin-tone line. Read the vectorscope, then use hue and sat moves to unify a sequence without running a beauty pass.
You do not need a beauty pass to unify a sequence of generated faces. You need the skin-tone line, and a small hue and sat move that puts every face on it. Models disagree on skin. The line does not.
This is not about why two models render different complexions. Skin tones jump when you mix video models is that diagnosis, and the fix there is a per-clip normalisation before any look. This is the instrument lesson: how to read the line, how generated faces typically miss it, and which two wheels actually close the gap without softening pores into plastic.
What the skin-tone line is
A vectorscope throws away luma and plots chroma. Angle is hue. Distance from the centre is saturation. Neutrals collapse to the middle. A saturated red spikes toward the R target.
The skin-tone line (also called the flesh line, and often drawn on top of the old NTSC I axis) runs between the red and yellow targets, roughly at 11 o'clock on a standard vectorscope graticule. It is a guide, not a law. Many software scopes default that line to 123°, the +I axis. That is a mark in a roughly ten-degree band, not a measurement of your talent's passport.
The useful fact underneath the folklore: human skin, across a wide range of complexions, is blood under scattering tissue. Melanin mostly changes how far out the cloud sits and how dark the face is on the waveform. It does not rotate the hue to green or to magenta on its own. Deeper skin often sits farther from the centre (more chroma) and lower on the waveform (less luma). It should still hug the same angle. A face that has rotated toward magenta or toward yellow-green has a colour error, a white-balance error, or a look that has been allowed to eat the talent.
Do not read the line on a monitor that has already adapted to the last clip. The scope does not adapt. That is the point.
How generated faces miss it
A camera has a matrix, a white point and a skin rendering you can profile. A sampler picks those as part of making the picture. Nothing in the clip is looking at the clip you already approved.
Typical misses, once you put a qualifier on the face and look only at that cloud:
| Where the cloud sits | What it usually is | What it is not |
|---|---|---|
| Rotated toward magenta / R | White balance too cool, or a model that over-paints blood | "They have pink undertones" as a creative choice you forgot to make |
| Rotated toward yellow / Yl | White balance too warm, or a model that went sallow / jaundiced | A daylight look. Daylight still puts skin on the line if the WB is right |
| Too far from centre | Oversaturation, high CFG-style punch baked into the render | Health. Skin that far out reads as makeup or as plastic |
| Collapsed on centre | Desat, a heavy print LUT, or a face that was generated grey-beige | "Cinematic" |
| Split into two lobes | Mixed light (practical plus daylight), or a qualifier that grabbed a red shirt | Two ethnicities. Two light sources |
A beauty pass tries to solve this by smoothing. Smoothing does not rotate hue. It will leave a magenta face smoother and still magenta, and it will take pores off a generated face that already had too few. If the sequence needs to match, match the cloud. If a still needs retouching, that is a different job, after the match.
Versely's video editor is where you assemble the sequence. It is not a vectorscope. preview: true gives a free 480p pass with a short per-user cooldown; the final export is charged once. Do the scope work in a colour tool that actually draws the line. The preview is for whether the cut reads, not for whether shot B's cheek is 8° off shot A's.
Hue and sat moves that unify a sequence
Work one clip at a time, face qualifier on, scopes up, look LUT off. A creative LUT will not fix mismatched shots. If a look is already on, bypass it until the faces agree, then put one look back on top of the match.
1. Qualify the skin, not the frame. Soft qualifier on the face and neck. Exclude lips, teeth, irises, wardrobe. A red jacket in the same vector as skin will yank the cloud and you will "correct" the jacket.
2. Set luma on the waveform first. Put the forehead or cheek in the same band shot to shot. Hue decisions on faces that differ by a stop are noise. You are not matching "beauty." You are matching exposure of the same region.
3. Rotate hue until the cloud sits on the line. Small numbers. A few degrees toward yellow if you are magenta; a few degrees toward red if you are sallow. Stop when the bulk of the cloud straddles the line, not when the most saturated pixel kisses it. Outliers are lips, ears and practical bounce. They are allowed to miss.
4. Bring sat in until the radius matches the master shot. Generated faces often sit too far out. Pull sat on the qualifier until the cloud's outer edge agrees with the shot you picked as master. Do not push sat on the dull clip to meet the loud one. Meet in the quieter place. Oversat skin is what makes a sequence look generated even after hue agrees.
5. Check a known neutral. A white shirt, a grey wall, a product card. If the face is on the line and the shirt is green, you rotated global hue rather than skin. Undo, tighten the qualifier, do it again.
6. Only then, look. One grade on the matched sequence. If the look pulls faces off the line, the look is too strong for generated skin, which already sits closer to plastic than photographed skin. Back the mix off. Contrast you can add with a gentle S-curve after; hue you cannot recover if the look has rotated it.
A sequence of six talking-head clips, three models, same room in the prompt: you will still do this six times. Prompt language does not lock a matrix. Matching generated footage to real camera footage is the same discipline when one of the shots is a real camera. The camera is not the authority because it is "real." The master shot you chose is the authority because you chose it.
Hex values in the prompt will not put a face on the line. Hex-locked colour is for a branded object that a model has been asked to hit as a number. Skin is not a hex. Treat it as a vector.
What you are not doing
You are not making everyone the same lightness. The line is hue. A darker face that sits on the line next to a lighter face that sits on the line is a matched sequence. Forcing both onto the same waveform band is a different, worse mistake.
You are not removing texture. If the cloud is on the line and the radius matches, stop. Sharpening, softening, pore synthesis and "beauty" are how generated skin loses the last of what made it look like a person. Grain and a little flare, if you need to break the AI sheen, come after, and they are not a hue tool; grain, flare and vignette is that pass.
You are not grading off a still of the face on your phone. Phones are wide-gamut, auto-white-balancing computers. The line is the check. The phone is the way the audience will see the mistake if you skip the check.
FAQ
Does darker skin sit somewhere else on the scope?
It usually sits farther from the centre and lower on the waveform. The angle should still be the line. If a darker face has rotated toward magenta or olive, that is still a balance error. Do not invent a second line for a second complexion.
What if the shot has no face?
Use a known neutral on the vectorscope (centre) and the RGB parade. If you have neither a face nor a neutral, you are matching by taste, which is how the jump survives. Generate a better take, or pick a different master.
Can a LUT put skin on the line for me?
A LUT multiplies whatever you feed it. If the cloud is 8° off, the LUT will still be 8° off, wearing a coat. Put skin on the line, then apply one look. If the look knocks faces off, the look is the problem.
Why not run a beauty pass and skip the scope?
Beauty does not rotate hue. It will leave mismatched faces smoother. On generated skin it will also erase the little texture the model gave you. Match first. Retouch, if you must, on the matched clip, as a separate decision.