Last month a creator sent me a folder of 40 stills pulled from a 2019 shoot. Every one of them was 640 pixels wide, compressed twice, and lifted out of a re-uploaded video. He wanted thumbnails from them. My first instinct was to say no. My second instinct, which is the one I acted on, was to run the whole folder through three different upscalers and see what actually survived.
Here is the short version of what I found. Upscaling has gotten genuinely good, and it still cannot save most bad stills. Those two things are both true, and the gap between them is where people waste a week.
What upscaling actually fixes
An upscaler is a guessing machine. It looks at the pixels you gave it, predicts what a higher resolution version of those pixels would look like, and paints that in. When the source has real information in it, the guess is close to right. When the source has lost the information, the guess is invention.
So the useful question is never "is this image low quality". It is "what kind of low quality is it".
Soft focus, mild noise, a slightly small but clean file, a screenshot that is sharp but only 800 pixels wide. Those upscale beautifully. The structure is intact. The model is filling in detail it can infer from what is already there, and on a thumbnail viewed at 320 by 180 in a feed, nobody will ever know.
Heavy compression blocking, motion blur, blown highlights, crushed shadows, and anything shot at high ISO in a dark room. Those do not come back. The upscaler will happily produce a crisp 4K version of a face that no longer looks like the person. I have watched a model turn a slightly blurry eye into a very sharp eye belonging to somebody else.
The rule I use now: upscaling recovers resolution, not information. If the detail was never captured, no model is going to hand it back to you.
The test I ran, and what the numbers actually mean
I took 12 stills from that folder and split them into three groups by failure type. Soft but clean, compression damaged, and motion blurred. I ran each through a general purpose diffusion upscaler, a dedicated face restoration model, and plain bicubic with a sharpening pass, which I included mostly as a control.
The soft but clean group came out usable across all three tools. The diffusion model was best, the face restorer was close, and even bicubic plus sharpening got two of the four to a point where I would have shipped them.
The compression damaged group was a coin flip. Two of four looked fine at thumbnail size and fell apart at full size, which for our purposes is acceptable, because a thumbnail is never viewed at full size. The other two developed that plastic, over-smoothed skin that reads as fake even in a small grid.
The motion blurred group failed completely. All three tools. Every image. The face restorer produced the most convincing failures, which makes it the most dangerous one to use without checking.
I want to be careful about how I frame the outcome side of this. We rebuilt six thumbnails for that channel using upscaled stills, and the ones that used clean-source upscales performed in the same range as the thumbnails we made from properly shot photos on the same channel. That is a small sample on one channel in one niche, and I would not tell you it generalises. What it tells me is narrower and more useful: a well-upscaled clean still does not carry a visible penalty. A badly upscaled one does, because viewers register "something is off" long before they can name it.
The face problem, which is the whole problem
Almost every thumbnail we build has a face in it. That is not a stylistic preference, it is just what works on most channels, and if you have read what a year of testing AI-generated thumbnails taught us you already know I am not romantic about hand-made for its own sake. I will use whatever tool produces the better image.
But faces are where upscalers earn their bad reputation. Human beings have absurdly good face detection. We notice when an eye is 3 percent wrong. We notice when skin texture has been replaced with a smooth gradient. We cannot always articulate it, and the thing we do instead is scroll past.
My working process for faces now:
- Upscale at the lowest strength that gets the job done. Most tools have a fidelity or creativity slider. Push it toward fidelity, not creativity.
- Check the eyes and teeth at 100 percent. Those are the two places models hallucinate first.
- Composite the upscaled face back over a cleanly scaled version of the original if the model changed too much. You often only need the sharpness, not the reinvention.
- If the person no longer looks like themselves, stop. Reshoot, use a different still, or build the thumbnail around something else.
That last one is the important one and it is the one people skip, because they have already spent 20 minutes in the tool and sunk cost is a powerful drug.
Cutouts make everything harder
Most of our thumbnails involve isolating a subject from a background. If you are working from a low quality still, the cutout stage is where the damage becomes visible, because you are now putting a soft, artifact-heavy edge directly against a clean flat background. Nothing hides.
The order matters more than people expect. Upscale first, then cut out. If you cut first and upscale second, the model has to guess at an edge that has already been hard-clipped, and you get that halo of invented pixels along the outline. Going the other way gives the masking tool more real edge information to work with, and the mask lands cleaner.
I wrote up the rest of that workflow in how we get photo cutouts that actually look clean, including the bits about hair and edge feathering that matter far more at small sizes than anybody admits.
The other thing that saves a soft cutout is separation that does not come from sharpness. A well-judged shadow or a subtle rim of contrast will read as "this subject is in front of that background" even when the edge itself is a little mushy. Just do not reach for the default settings, because the way drop shadows quietly date a thumbnail is one of the fastest ways to make a 2026 thumbnail look like a 2014 one.
When the honest answer is "use a different image"
There is a version of this article that ends with "here are the five best upscalers, go forth". I am not going to write that one, because the tool matters much less than the decision of whether to upscale at all.
If a creator has 200 stills and 15 of them are clean, we use those 15 and design around them. That is almost always better than heroically rescuing a bad frame. And if there are genuinely no usable stills, the real options are a quick phone reshoot against a plain wall, which takes 10 minutes and beats any model, or a composition that does not need the face at all.
The failure mode I see most often is a creator who reaches for stock instead, which solves the resolution problem and creates a worse one. A stock face on a personal channel is a mismatch that viewers feel immediately, and I have gone on about why stock photography makes every channel look like every other channel at some length. A slightly soft photo of the actual host beats a pristine photo of a stranger. Every time.
What this means if you are doing it yourself
Upscaling is now a normal part of a thumbnail workflow rather than a specialist trick, and the tools are cheap or free. That is real progress. The catch is that the tools are also confident, and confidence is not accuracy. A model will never tell you it failed. It will just hand you a sharp wrong image and let you find out in the analytics three weeks later.
So build the check into the process. Look at the result at actual thumbnail size, in a feed, next to other thumbnails. That is the only size that matters. Then look at it at 100 percent to catch the hallucinations. If it passes both, ship it.
And be realistic about the time. A clean upscale takes two minutes. A rescue attempt on a genuinely damaged still takes 40 minutes and usually fails. Knowing which one you are looking at inside the first 30 seconds is the actual skill.
How we handle it at ClickTheory
Every thumbnail we make is custom for that specific video. We do not run templates, and we test styles on our own channels before we put them in front of a client, which is also how I ended up doing the upscaler comparison in the first place. Source repair is just part of the job. It is included in the work, not billed as an extra step.
Our pack is $250 for 8 custom thumbnails, which works out to about $31 each, one per video, with unlimited revisions and a dedicated designer. Title consultation is included, because a thumbnail and a title are one unit and pretending otherwise is how good images get wasted. Turnaround is 48 to 72 hours per thumbnail. If you want to see what comes out the other end, our work is the honest version.
For creators weighing this against doing it in-house with a template tool, I would read the case for and against Canva first, and then how a pack compares to an unlimited design subscription if you are trying to work out what you actually need. If your show is a podcast and the stills are coming out of a multicam recording, my other studio edits video podcasts and cuts clips, and honestly, fixing the export settings upstream will do more for your thumbnails than any upscaler will.
The tools got good. They did not get magic. Feed them something real and they will make it bigger. Feed them nothing and they will make nothing bigger, very sharply.