October 12, 20266 min read

Making YouTube Thumbnails with AI, Tested Model by Model

We gave ChatGPT and Gemini the same three Korean thumbnail briefs and recorded the first result of each — no regenerations. Hangul never broke. Something else did. Then we ran the same topic through a channel-fitted pipeline to see what actually changes.

The first thing creators ask about AI thumbnails is whether the text comes out broken — in Korean, whether Hangul glyphs survive. We wanted the current answer, not the 2024 answer, so we ran the test ourselves: the same briefs, the same prompt, first results only.

The short version: the headline text was perfect in all six images. The failures moved somewhere subtler — text we never asked for.

How we measured

Three briefs, chosen to stress different things: a finance video ("월급 200으로 1억 모은 방법", on-image text "월 200 → 1억"), a diet video ("3일 만에 -5kg, 운동 없이", text "운동 없이 -5kg"), and a tech video ("ChatGPT 새 기능, 이건 진짜 미쳤다", text "진짜 미쳤다"). Numbers, arrows, and syllable-final consonants included on purpose.

Every model got the identical Korean prompt: "유튜브 썸네일을 만들어줘. 16:9. 영상 제목: {title}. 이미지에 넣을 글자는 정확히 '{text}' — 다른 글자는 넣지 마. 클릭하고 싶어지는 구도로." (Make a YouTube thumbnail, 16:9. The only text in the image is exactly this — no other letters. Compose it so people want to click.)

ChatGPT (GPT Image, web UI) and Gemini via AI Studio — the full Gemini model (Nano Banana 2.1) required a paid API key to run in AI Studio, so the free tier, Nano Banana 2 Lite, is what we measured. That gate is itself a finding.
First result only. No regenerating until it looks good — that would measure our patience, not the models.
Run on 2026-10-12. Small sample, so read this as a field report, not a benchmark.

Result 1 — big Hangul doesn't break anymore

All six headlines rendered letter-perfect, including the 받침-heavy syllables and the mixed number strings. Two years ago this test would have been a gallery of invented glyphs; today, short and large Korean text is a solved problem on both frontier models.

With one condition: small text is still decorative. Inside Gemini's tech thumbnail, the little English paragraph on the computer screen reads as plausible-looking gibberish up close. The big text is real; the fine print is set dressing.

Practical rule: keep on-image headlines short and large, in any tool.
If a design needs small text, add it yourself afterwards — don't let the model typeset it.

Result 2 — uninvited text shows up

Three of the six images contained text we never asked for, despite the prompt explicitly saying no other letters. Not broken text — plausible text, in places a viewer might not check:

ChatGPT, finance brief: a payslip prop printed with "급여 2,000,000" (salary 2,000,000).
Gemini, finance brief: "1억" repeated inside a gold coin, and "기재부" printed on a book prop.
Gemini, tech brief: an entire invented feature panel — a "NEW FEATURE" header, logos, and feature labels.

This is the failure mode that matters in practice. A garbled glyph is obvious; a believable extra word on a prop ships quietly and costs trust. Whatever tool you use, zoom into the props and backgrounds before you publish.

Result 3 — similar speed, different friction

Generation time was a wash: ChatGPT took roughly 20–55 seconds per image, Gemini Lite 25–60. The real difference was before the first image: ChatGPT generated immediately, while Gemini's full model sat behind an API-key gate in AI Studio and only the Lite tier ran free. For a creator who wants one thumbnail now, the smoothest path was ChatGPT's.

What changes with a fitted pipeline

We then gave AutoKliq — our own tool, so weigh that — a single input: a topic ("ChatGPT 새 기능 총정리") on a real channel it could read. This is a different shape of test by design: raw models take a prompt, a pipeline takes a channel. In about eight and a half minutes it returned six planned concepts, each with its own headline, composition, and stated reason it should work.

Three differences held up under inspection. First, the headlines never repeated the title — they were curiosity lines in the channel's voice ("그거 이제 동료가 합니다" — "your coworker does that now"), which is the packaging rule raw prompts don't know. Second, all six renders had letter-perfect on-image text and zero uninvited text: screens inside the images are rendered as abstract bars, because the pipeline refuses to draw text it cannot verify, and a transcription check compares drawn text against the intended headline after rendering. Third, the face was the channel's actual creator, consistent across concepts, where the raw models invented strangers — they had no reference, which is exactly the point: the pipeline starts from your channel, the prompt box starts from zero.

The honest trade-offs: time-to-first-image is slower (minutes of analysis before any render), and the image model underneath is the same class as the raw tools. What a pipeline adds is not a better brush — it's the planning before and the verification after.

So how should you make thumbnails with AI

None of this is a reason to abandon the raw models. It's a reason to know which step each tool is actually for:

For one quick thumbnail, a raw model is genuinely fine now — big Korean text included. Budget two minutes for inspection: zoom into props, backgrounds, and any small text.
Never ship small text the model typeset. Add fine print yourself.
If every video needs channel-fitted planning — headlines that complement the title instead of repeating it, your face, your style — and you want the inspection automated, that's the pipeline's job.
Either way, the durable rules didn't move: headline under four words, complementary to the title, one idea per image.

Limits of this test

Three briefs, two raw models, one pipeline run, one day. Models change monthly; treat the specifics as a dated snapshot and the failure modes — uninvited text, fake fine print — as the parts likely to outlast it.