One video, four AI tools
A lot of YouTube isn't in English, and a thumbnail tool shouldn't need it to be. ThumbnailPeak's four AI tools read your video in whatever language it's in. To show that, we took one Russian-language video through all four, start to finish, without translating a word.
The video is "РЕАЛ МАДРИД - НОРМ?", roughly "Real Madrid: doing OK?", from the fan channel Мадридист / Юра Калинов (Madridista / Yura Kalinov). It was published on October 8, a month into the season, and in it Yura and six other fan-bloggers grade José Mourinho's start at Real Madrid. By October 12 it had about 19,000 views.
Here's the loop:
- Generate made a thumbnail from the YouTube link.
- Critique graded it and listed what to fix.
- Improve made those fixes in one click.
- Compare put the two versions head to head and picked one.
All four results are public, and the links are at the end of this post.
Step 1: Generate, from a Russian link
![]()
The channel's own thumbnail sells the month: "МЕСЯЦ МОУРИНЬЮ" ("Mourinho's month"), Mourinho at the centre, and the guests' channel logos around him.
Generate went another way. The video's description, in Russian, promises "an honest conversation about the Vinícius and Mbappé problem, which has only got worse," and that's the angle Generate picked. It shows the two forwards, split down the middle, under one question: "НЕ СРАБОТАЛИСЬ?" The verb is сработаться, two people learning to work well together. The closest English is "Didn't they click?"
The title and description went in in Russian, and the headline came out in Russian, because this is a Russian-language channel. Every concept Generate shows comes with a full critique score, graded against the video's real title. This one got 85, Strong.
Is it better than the creator's own thumbnail? That's not what this run set out to show. A link run gives you angles you haven't tried. If you want the head-to-head, the Compare vs current thumbnail button on the card puts a concept against the thumbnail the video has now.
Step 2: Critique, in Cyrillic
Next we ran a full critique on the generated thumbnail. One thing to know before the numbers: the title box held a placeholder, the single word "nice." That turned out to be a useful test, because the critique noticed.
It came back 78 overall: visual 90, title fit 8, "Strong," low confidence.
It read the Russian. Legibility was graded at the size YouTube shows in its sidebar: "The yellow-and-white Russian headline remains readable at 168×94." Curiosity value got 3 out of 4, because "the question headline and split image suggest a disagreement or failed partnership between the two players." That is what "НЕ СРАБОТАЛИСЬ?" means, read off the image and explained in English.
It knew who was in the picture. Hierarchy, subject clarity and contrast all got the top grade, 4: "Both footballers are large, unobstructed, and recognizable at feed size," with "thick black outlines against a dark stadium background." Its fit notes name both players, Vinícius Júnior and Mbappé, without being told.
It found the picture's weak spot. Distinctiveness got 2: "The split-screen footballer comparison uses a common sports-thumbnail format without a distinctive visual idea."
And it caught the title. Topic match got 0: "The title “nice” gives no indication that the video is about football, Vinícius Júnior, or Mbappé." That is where the fit score of 8 comes from. Remember that Generate scored the same picture 85 against the real title.
The fix list had two items:
- High priority: align the promise with the content. "Rename the video to clearly state the relationship or conflict shown by the thumbnail, such as “Did Mbappé and Vinícius Get Along?”" The reports are written in English, so the suggested title came back in English, but the idea carries straight over to Russian.
- Medium priority: stand out from the genre. "Add one distinctive, topic-specific visual cue such as a failed-play moment, tactical graphic, or clear disagreement detail."
Step 3: Improve, in one click
![]()
Improve works from the critique's fix list. No image edit can rename a video, so the change it made answers the second fix. Behind the same two players it put a Real Madrid flag, the club crest on the dividing line, and the Bernabéu's floodlit stands, with "SANTIAGO BERNABÉU" lettered on the stand. The players, their poses and the Russian headline came through untouched, letter for letter.
We critiqued the result with the same placeholder title. It came back 84 overall: visual 93, title fit 33, "Strong," and this time high confidence.
![]()
Distinctiveness went from 2 to 3. "The split-player layout and Real Madrid branding are easy to recognize."
The fit went up under the same one-word title, because the picture now names the club on its own. Topic match: "the thumbnail clearly presents a Real Madrid player comparison." Promise consistency, 2: "Nothing in the short title directly contradicts the thumbnail, but it does not support its question."
Not everything moved. Trust & polish stayed at 3: "The oversized headline and bright center divider make the image feel heavily assembled but still professional." Curiosity value stayed at 3 as well: "the question gives little specific context."
What's left to fix is the title. All three remaining fixes point at "nice," starting with "Replace NICE with a specific title naming the players or the Real Madrid situation." The fix list had nothing left to say about the picture.
Step 4: Compare, head to head
Compare puts two thumbnails side by side against the same title and asks which one a viewer is likelier to click. It picked the improved version, by a slight margin, at low confidence:
Its reasoning: "A is the stronger click choice because it makes the Real Madrid setting clearer at feed size. The stadium, large club crests, and two players all remain visible, while B's blurred background removes much of that context."
It also gave the original its due. On background clarity it went with the generated version: "B's soft background gives the faces more separation; A's sharp crowd, lights, and signage add some visual competition." You can see both points at the size most people will meet this video:
![]()
A slight win at low confidence means both thumbnails are shippable and the improved one is the better bet. If you want to settle it with real viewers, YouTube's Test & Compare runs up to three thumbnails on the same video.
What "any language" means here
- Your words can be in any language. Generate read a Russian title and description. The critique and the comparison read a Cyrillic headline off the image and graded its legibility at feed size, the same way they grade English.
- Generate writes for your audience. Here that meant a Russian headline for a Russian audience. If you want something else on the image, such as English text on a Russian video, say so in the guidance box. Your guidance outranks its defaults.
- Improve leaves your words alone. The headline came through letter for letter, while everything around it changed.
- The reports are in English. Grades, evidence and fixes come back in English, so a suggested title is yours to put into your own language.
Three things to steal
- Put the conflict in your description. Generate's angle came from one sentence of it: "the Vinícius and Mbappé problem, which has only got worse." If your description says what the video is really about, the thumbnail can say it too.
- Show the specific. A crest and a stadium were worth a level of distinctiveness. They also lifted the fit score under a one-word title, because they told viewers which club the video is about.
- Never ship a placeholder title. Every critique in this loop put the title first. The same picture scored 85 against the real title and 78 against "nice."
Run the loop on your own video
Start in Generate. Paste a YouTube link in any language, public or unlisted, or describe a video you haven't uploaded yet. From the card, Critique gives you the full report, and Improve this one sends the card straight to Improve, starting from its own critique. Compare puts any two versions head to head. A free account covers one full loop every month: one generation, two critiques, one improve and one comparison.
Every result in this post is public:
- The generated thumbnail
- The critique of the generated thumbnail
- The critique of the improved thumbnail
- The comparison
For more on what Generate can do, from a link or from a paragraph, see How to Get Good AI YouTube Thumbnails From a Link or a Paragraph.
The video and its current thumbnail belong to Мадридист / Юра Калинов and are shown here for commentary and education. The generated and improved thumbnails were made with ThumbnailPeak for this post and are not the channel's. English glosses of the Russian are our own.