I used to grade thumbnails the same way I grade full sequences: one global adjustment, maybe a curve, done. It never worked. A thumbnail has to read in half a second at 200 pixels wide on someone's phone, and flat global edits just don't survive that kind of compression. What actually fixed it was leaning hard on Lightroom Classic's subject and sky masking, which by late 2025 has gotten good enough that I barely touch the brush tool anymore.
Split the frame before you touch color
My process now starts with three masks on almost every still: one for the subject (usually a person), one for the sky or background, and one radial mask centered on whatever I want the eye to land on first. I build these before I do any color work at all. Once they exist, everything downstream is faster because I'm nudging three separate exposure and contrast values instead of fighting one slider that has to serve the whole image.
The subject mask gets a small contrast boost and a touch of clarity, nothing dramatic. The background mask usually gets pushed a stop darker and slightly desaturated. That's the whole trick, honestly. Darken everything that isn't the subject and the subject looks lit even when it wasn't.
Where people overdo it
The failure mode I see constantly, including in my own early attempts, is cranking the subject mask too hard. Skin starts looking plastic, edges get a faint glow where the AI selection wasn't perfectly clean, and the whole image reads as fake from three feet away. I keep every subject adjustment under +15 on any given slider. If it needs more than that, the original exposure was wrong and no mask will fix it properly.
- Subject mask: +10 contrast, +8 clarity, nothing else
- Background mask: -0.6 exposure, -15 saturation
- Radial mask on the focal point: +5 exposure, feathered to 90
That's genuinely the whole recipe for eighty percent of my thumbnails. It sounds too simple to matter, but the eye is lazy and it goes exactly where the contrast tells it to.
A note on AI select accuracy
Lightroom's automatic subject detection still trips on hair against busy backgrounds and on hands holding reflective objects, cameras especially. When it messes up, I don't fight the AI mask. I duplicate it, subtract the bad area manually, and move on. Trying to get the algorithm to behave perfectly wastes more time than a fifteen-second manual fix.
None of this is groundbreaking. What surprised me is how much a thumbnail's click-through rate depends on this kind of boring, repeatable separation between subject and background rather than on any single dramatic color grade. I'll take consistent over clever most days.