Precision instrument for YouTube thumbnails

Fix the thumbnails costing you views

Rexels reads your channel, finds the videos that deserve better, remakes their thumbnails from your own video, and ranks the result against your own history. We measure it. We prove it.

I tried the
1998 GP
current22
beats 22% of your uploads
The fastest lap ever
remade78
predicted to beat 78%conf. high
Ranked against your own history/Not an internet average/Read-only channel access
Fig. 01 — one video through the loop
90-day
analytics window
9
objective signals per thumbnail
1,024-dim
visual fingerprint
0
invented scores
The obvious question

Gemini already makes images

So does ChatGPT. So do we, with the same class of models. The image was never the hard part. Knowing which thumbnail is costing you views, what to replace it with, and whether it worked, that is the hard part.

A generator gives you

  • An image, usually a good one.
  • A 0 to 100 score invented on the spot.
  • The same advice it gives every other channel.
  • No idea which of your videos needs it.

The loop gives you

  • Which of your videos are underperforming, ranked by what a remake is worth.
  • A remake in your visual language, with your real face.
  • A prediction calibrated to what has won and lost on your channel.
  • The measured before and after, plus how often our calls were right.

A model can paint. It cannot read your analytics, rank your history, or own a mistake.

The loop, end to end

Diagnose. Remake.
Predict. Prove.

01

Connect, read-only

Link YouTube in one click. Rexels pulls 90 days of real analytics: impressions, click-through, watch time. Read-only scopes, so nothing on your channel can be changed by us.

02

Diagnose

Every thumbnail is scored against your channel's own baseline, long-form and Shorts in separate cohorts. High-impression, low-CTR videos surface first: the ones a remake can actually help.

03

Remake

Regenerate from your actual video, in your visual language, with your real face. GPT-Image, FLUX, and Gemini do the painting; the brief comes from your data.

04

Predict

Candidates are ranked against the thumbnails that already won and lost on your channel, with the reasons spelled out. Too little history to call it? It says so instead of inventing a number.

05

Prove

Publish and watch the real numbers. Rexels keeps score on itself: it checks its calls against your realized click-through and shows the running accuracy on your channel.

Method

Measured, not vibes

Every thumbnail score you have seen was a language model's opinion. Ours is a measurement. Each candidate goes through the same instruments before anything is ranked.

Where the eye lands

Attention models trained on human gaze pick out the focal region, and how much of the image is fighting it.

Legibility at phone size

Every word of overlay text is read with its pixel geometry, then judged at the size a phone feed renders.

Clutter

A vision-science busyness measure on the shrunk image, because fine detail turns to mush at 120 pixels tall.

Subject presence

How much of the frame your face or subject actually owns, counted in pixels rather than guessed at.

Your visual fingerprint

A 1,024-number signature of each image, compared against the thumbnails that already won and lost for you.

Candidate B, measured
Faces detected
1
Subject share of frame
41%
Focal region
tight, centered
Overlay text
3 words, legible
Clutter
0.34 / 1
Similarity to winners
0.81
Predicted to beat 74% of your uploadsconf. high
Fig. 02 — one candidate's readout

Then Rexels checks its own work. The ranking is validated against your realized click-through, out of sample, and the running hit rate shows in the studio. When your history is too thin to call, you get not enough data: the one answer a vanity score will never give you.

For small channels

Too small to A/B test? That is exactly who this is for

Live testing needs thousands of impressions per variant. Below that, YouTube's own Test & Compare comes back Inconclusive, and so does every paid split tester. The tools built to fix low click-through do not work for the channels that need them most.

Rexels predicts before you publish, from your own history plus the visual evidence. A channel with twenty videos gets a calibrated call, not a coin flip. And once you are big enough to test live, Rexels hands its best candidate to YouTube's referee instead of pretending to be one.

Personas
In the studio now · free to create

Your face. Actually your face

In one late-2025 survey, nearly half of creators who tried AI thumbnails had already quit, and the reason was almost always the same: the plastic look. An audience scrolls straight past a face that is almost you.

01

Many angles, not three

Keep up to 14 reference photos per persona, and every generation is conditioned on the strongest set of them. Tools that cap you at three are exactly why their faces drift into someone almost you.

02

Preserved, not pasted

Your likeness carried inside the generation, reference-conditioned, not swapped on afterwards. That is the difference between you and almost-you.

03

Hybrid by design

AI rebuilds the scene, lighting, and text; the face stays yours. Rexels already remakes thumbnails from your own video frames, and personas complete that principle.

If it doesn't look like you, it doesn't leave the studio.

The field

Where Rexels sits

Every tool here is good at its job. None of them close the loop from your analytics to a remade thumbnail to a measured result.

CapabilityRexelsAI generatorsChannel coachesA/B testersChat AIs
Generates the thumbnailsome
Reads your analytics, read-onlyduring tests
Tells you which videos to fixgeneric
Prediction calibrated to your channel
Works without live-test traffic
Real face, many reference photosbetaa fewa few
Measures real before-and-after CTRbeta
Reports its own accuracy
Admits when data is thin

AI generators: Pikzels, GoThumbnails. Channel coaches: vidIQ, TubeBuddy. A/B testers: ThumbnailTest, YouTube Test & Compare. Chat AIs: ChatGPT, Gemini. Public information as of mid-2026; if anything is out of date, tell us and we will fix it.

Named comparisons: vs Pikzels · vs vidIQ · vs ThumbnailTest

Pricing

Priced like a tool, not a tax

A remade thumbnail that recovers one underperforming video pays for the month. Generation is metered in credits, so heavy months cost more and quiet months don't.

To start

Free

$0
  • 100 credits on signup
  • Every AI model
  • General best-practice ranking
  • AI titles
Start free
Growing channels

Creator

$29/mo
  • 600 credits monthly
  • Connect one channel, read-only
  • Thumbnail health scores
  • Predictions calibrated to your channel
  • Real-face persona (beta)
Start free
Most popular
The full loop

Pro

$49/mo
  • 1,200 credits monthly
  • Everything in Creator
  • Underperformer queue
  • Before-and-after lift tracking (beta)
  • Prediction accuracy self-audit
  • Premium models
  • Priority support
Start free
Teams, multi-channel

Studio

$129/mo
  • 3,000 credits monthly
  • Everything in Pro
  • Multiple channels
  • Team seats
  • Bulk remake across your catalog
Start free

Every plan starts free today. Paid tiers open with the public beta; annual pricing saves about a third.

7-day money-back guarantee on every paid plan. Cancel is a button, not a support ticket.

Questions

Asked, answered

Ready to close the loop?

Connect your channel, read-only. See your worst thumbnails in minutes. Remake the first one today.

Connect your channel

No card required · 100 free credits