Google's own image-research team has been open about how tools like this actually work. They built a model called NIMA, short for Neural Image Assessment, trained on roughly a quarter million photos and how they scored across nearly a thousand different online photo contests.
Then they used that score as the literal training target for automatic photo enhancement. Google's own writeup on the project says plainly that results can be improved "by contrast adjustments directed by the NIMA score." Read that again. The goal was never "make this photo look like what you actually meant."
The goal was "make this photo score higher against a popularity model trained on internet photo contests." Punchier contrast and richer saturation score better on average, across an average photo, for an average viewer. So that's what you get. Every time, on every photo, whether it needed it or not.
There's a second reason the results look fried, and it's a plumbing problem, not a taste problem. Most of these tools operate on the final, already-flattened image, the version with all its contrast and color already baked in.
Push contrast and saturation hard on a file like that and there's nowhere for the extra information to go. Highlights clip to pure white. Shadows clump into pure black. Detail that was actually sitting in that file gets thrown away, permanently, the moment you save.
Imaging researchers have a name for the visual result of this kind of aggressive processing: "radioactive" images, over-lit and over-sharpened, with a faint glow around anything with real contrast. You've seen it. It's the sunset photo where the sun has a cartoon outline burned around it.
How to Tell a Real Grade From a Fried One
The Three Separate Moves
Working colorists split this job into three separate moves, and the split is the whole point. Correction is the technical fix, the unglamorous work of getting exposure and white balance neutral before anything creative happens.
Grading is the deliberate choice layered on top of that, made shot by shot, on purpose. A filter is neither of those things. It's one fixed adjustment slapped over the top of everything, applied identically whether a given photo needs it or not.
The color science standard used across professional film and video (ACES) treats these as genuinely separate operations, and for a real reason: keeping them apart is what lets someone change their mind about the creative look later without redoing the technical work underneath it. Collapse all three into a single auto-enhance button and that separation disappears completely. You get a fixed look bolted straight onto an uncorrected image. No correction step. No creative intent. Just a preset wearing a decision's clothes.
What Real Grades Actually Look Like
Here's the tell, once you know to look for it. A real grade treats every part of an image differently on purpose. Skin tones stay close to natural even when the sky goes moody and dramatic behind them. Shadows keep some texture instead of collapsing into flat ink.
A one-click filter can't manage that, because it isn't actually looking at your photo. It's applying the same move to every photo it's ever seen. That's also why it falls apart across a whole shoot.
Run twenty photos from one session through an auto-enhance tool and you'll get twenty slightly different guesses at "vivid," because the tool is reacting to each frame in isolation instead of holding one consistent look across the set. A colorist matching shots by eye catches that instantly. An algorithm scoring each frame on its own has no idea it's even a problem.
Flat Footage Makes All of This Worse
If you shoot video, you've probably met a meaner version of this same issue. Flat or "log" camera profiles are recorded deliberately dull and gray on purpose, to protect highlight and shadow detail for later. Think of it as a digital negative, not a finished image.
It needs a specific technical step to turn back into something that looks normal before anyone touches the creative side at all. Feed flat log footage straight into an auto-enhance tool and it reads "gray and boring," then stomps on contrast and saturation to compensate, with zero awareness that the file was never supposed to look right in the first place.
The result is exactly what you'd expect. Blocked-up shadows. Banding across a sky that was supposed to be a smooth gradient. Skin that goes plastic in precisely the way experienced colorists warn beginners about, constantly, because it's such a common mistake.
Control Is the Whole Point, Not an Upgrade
None of this makes AI the enemy. Good AI-assisted tools are genuinely useful. They just need to leave the actual decisions to you instead of making them invisibly on your behalf. The real question is control, not intelligence.
Can you see what the tool changed and adjust it? Can you pull the look from one photo you love and apply that exact look, on purpose, across the other forty photos from the same shoot, instead of getting forty different guesses? Can you check your work against a real measurement instead of your monitor and a feeling?
That's the actual bar. It was never about fighting AI. It's about refusing to hand off judgment calls that were always supposed to be yours. We built Color Engine's reference color match around exactly that idea: pick one photo or frame with the look you actually want, and match everything else to it deliberately, with real scopes so you can see what's happening to your highlights instead of guessing. If you've been burned by the one-click version of "AI color," that's not a personal failing. It's a sign you were paying attention. Go see what real control actually looks like, and stop letting a popularity contest from a stock photo site decide what your work is supposed to look like.
