I spent about four hours last Tuesday trying to get Midjourney to produce one thumbnail. Not a mood board, not a background plate, one finished thumbnail for a finance video about interest rates. I gave up at prompt sixty something and opened Photoshop, and the thing that finally worked took eleven minutes.
That is not a story about AI being bad. Midjourney is genuinely excellent at what it does. It is a story about the gap between "makes a beautiful image" and "makes a thumbnail that earns a click at 168 pixels wide." Those are different jobs, and the second one is mostly not a rendering problem.
We test styles on our own channels before we put them anywhere near client work, so I have run a lot of these experiments with nobody's growth on the line. Here is the honest list of what generative image tools still cannot do for thumbnails, why each gap exists, and where they have actually become indispensable.
It cannot design for the size the thumbnail is actually seen at
Midjourney composes for the frame it is given. It fills the canvas, balances it, adds detail everywhere because detail looks good in a square preview on a large screen.
A YouTube thumbnail on a phone home feed is roughly the size of a postage stamp. On a TV it is bigger but viewed from three metres away. The design constraint is not "make this beautiful," it is "make this legible in about 0.4 seconds at a size where three quarters of the detail is invisible."
Every generated thumbnail I have tested fails the squint test on the first pass. The subject sits at mid scale instead of dominating. Background elements compete. The value range is compressed into a pleasant middle grey mush because that is what looks tasteful at full size and what disappears entirely at thumbnail size.
You can fix this. You crop hard, you push the subject up 30 percent, you crush the blacks, you rebuild separation with a rim light that was not in the generation. At that point you are designing, and the AI gave you a starting texture rather than a thumbnail.
The model optimises for how the image looks when you are looking at it. Thumbnails have to work when nobody is looking at them yet. Those are opposite problems.
It cannot keep a face consistent across a channel
This is the single biggest practical blocker and the one people underestimate.
A channel is a brand, and on YouTube the brand is very often a face. Same person, same lighting logic, same expression vocabulary, recognisable in a sidebar next to eleven competitors. Character reference features have improved a lot, and for a stylised mascot they are close to usable. For a real creator's real face they are still not there.
The failure is subtle, which makes it worse. The face is 94 percent right. The jaw is slightly narrower, the eyes sit a hair further apart, the skin has that faintly plastic evenness. Viewers do not consciously notice. They just do not get the flicker of recognition that makes them stop scrolling, and you never see that in the analytics because the click that did not happen leaves no trace.
We wrote about the broader version of this problem in why synthetic faces cost you trust, and the short version is that the uncanny cost lands on the channel, not on the image. Faces also carry an outsized share of visual attention, which is exactly why small errors there are expensive.
For our packs we still cut the creator's real face out of real footage or a real shoot. Every time. It is slower and it is not close.
It cannot read text, and it still mostly cannot render it
Text rendering has improved from "gibberish" to "usually right on short words." That sounds like progress and it is, but thumbnail text is not a rendering task at all.
Thumbnail text is a hierarchy decision. Which three words survive. What weight, what tracking, what stroke, whether the word overlaps the subject or sits clear of it, whether it repeats the title or completes it. Getting "PROFIT" to appear in the image is the trivial part. Deciding that the word should be "PROFIT" and not "I LOST 40K" is the entire job.
Generative models have no access to the input that decision needs: the title, the first fifteen seconds of the video, the channel's usual promise, what the last six thumbnails already said. That context lives in a conversation with the creator. It is why title consultation is bundled into our pack rather than sold separately, and why briefing a designer properly changes the output more than any tool choice does.
Also, practically: generated text is baked into pixels. You cannot nudge the kerning, swap a word for the A/B variant, or fix it in two minutes when the creator changes the title an hour before publish. Layered files matter more than people expect.
It cannot tell you whether the thumbnail worked
This one is obvious when stated and ignored constantly.
Midjourney has no feedback loop with your audience. It produces an image that satisfies an aesthetic prior learned from a very large pile of pictures, most of which were never thumbnails and none of which were your thumbnails.
We ran a fairly clean test on one of our own channels in the education space, sitting somewhere in the mid five figures of subscribers. Same video, same title, two thumbnails: one fully generated and polished, one shot and composited the usual way. The generated one lost, but only by about 1.1 points of CTR, which is smaller than I expected and honestly a point in the model's favour. What was more interesting was the second order effect. The generated version pulled a slightly worse average view duration, which suggests it was attracting a marginally less qualified click.
I would not generalise that to your channel. It was one video, one niche, one audience. But it matches something I keep seeing: the interesting question is not which image wins the click, it is which image wins the click from someone who stays. That tension is the whole subject of click through rate versus watch time, and it is the reason we treat a raw CTR bump with some suspicion.
Testing costs us fifteen dollars for an A/B variant. That is a rounding error against being wrong for a month.
What it is genuinely great at
I am not writing an anti AI piece. Roughly half of what we ship now touches a generative tool somewhere in the pipeline, and I would not go back.
Backgrounds are the clearest win. If I need a moody server room, a stylised cityscape, an abstract gradient with real texture in it, I generate it in ninety seconds instead of hunting stock for forty minutes. Nobody is examining the background. It just needs to be a surface with the right temperature and enough noise to feel real.
Props and objects, same story. A floating credit card, a cracked trophy, a specific kind of chart on a specific kind of paper. These are composited small, they carry no identity, and small errors are invisible.
Concept exploration is the underrated one. When a creator is not sure what the video even looks like visually, twelve fast generations move that conversation faster than any amount of describing.
And the tooling around the generation has quietly become better than the generation itself. Modern cutout and matting tools do more real work in our process than any prompt does. We went through that whole category in our comparison of background removal tools and the gap between the good ones and the free ones is larger than you would guess.
The composite is the actual craft
Here is the mental model I have landed on after a year of this.
A thumbnail is not one image. It is usually four to nine elements stacked with intent: a real subject cut cleanly, a background that sits behind it in a different value range, a light or glow that separates the two, text placed to lead the eye, and one or two props doing narrative work. The skill is in the stacking, the contrast decisions, the ruthless removal of anything that does not earn its place.
Midjourney makes one image. It makes a very good one image. But a single flat generated frame gives you nothing to push, no separation to control, no layer to move when the creator says the face should be bigger. You lose the exact controls that thumbnail design consists of.
So we use it as a supplier of elements, not as a designer. That is not a downgrade. It is just an accurate description of what the tool is.
I collected what a full year of running these experiments actually changed about our process in a year of AI in the thumbnail pipeline, including the two workflows we abandoned entirely.
What this means for what a thumbnail costs
The honest question underneath all of this is usually financial. If a tool can generate an image for pennies, why does a custom thumbnail cost real money.
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, and 48 to 72 hours per thumbnail. Roughly none of that price is the rendering. It is the decision about what the thumbnail should say, the crop, the contrast, the three rounds where we make the face read at phone size, and the willingness to throw out a nice image because it is not a working one.
Generative tools have made the picture making part cheaper. They have not touched the judgement part, and the judgement part was always the expensive bit. If you want the full breakdown of where the money goes, what thumbnails should cost lays it out with fewer opinions than this post has. You can also just look at our work and decide whether the difference is visible to you. It should be.
The same logic applies elsewhere in the content stack. AI has not replaced podcast editing either, though it has made transcription and rough cuts vastly faster, which is why our sister service edits video podcasts with humans still making the cut decisions.
Where I expect this to break
I will happily be wrong about most of this within eighteen months. Face consistency is the one I expect to fall first, because the money pointed at it is enormous and the problem is well defined. Text rendering will keep improving. Native layered output would be the real unlock, and somebody will ship it.
What I do not expect to be solved is the part that was never a model problem. Knowing that this creator's audience responds to restraint while that one's responds to chaos. Knowing that the video is actually about regret and not about money. Deciding which three words are the ones.
That comes from watching the channel, talking to the person, and being wrong a few times in public. No prompt gets you there.
If you are already generating your own and just want someone to tell you honestly whether they are working, send me two and I will tell you. That is a five minute favour, not a pitch.