Making Blog Thumbnails with AI Image Generation

Confessions of Someone Who Can't Even Draw a Rectangle in Figma

Running a blog taught me an uncomfortable truth: the same post gets way more clicks when it has a thumbnail. But I can't design. Like, genuinely cannot. I open Figma and have no idea what to do.

So I turned to AI image generation. Over the course of a month, I rotated between Midjourney, DALL-E 3, and Stable Diffusion to create thirty blog thumbnails.

Midjourney: Beautiful but Doesn't Listen

Images from Midjourney v6 are undeniably gorgeous. Feed it "minimalist blog thumbnail about React hooks, clean geometric shapes, blue and purple gradient" and you get something gallery-worthy.

The problem is getting the composition I want. When I say "leave empty space on the left for text," it doesn't really understand. Maybe two or three out of ten attempts produce a usable layout. The rest are so filled with imagery that overlaying text destroys readability.

DALL-E 3: Actually Follows Instructions

DALL-E 3 follows prompts relatively accurately. "Leave the left third empty, with an abstract code editor on the right" comes out as intended. The art quality is slightly below Midjourney, but for blog thumbnail purposes, this was the more practical choice.

This Is Where I Found My Pattern

After plenty of trial and error, I settled on a workflow: generate a background image with DALL-E 3, add text and logo in Figma, then apply a brand color overlay.

I also templatized my prompts. "Abstract [topic], minimalist style, [color] palette, soft gradients, clean composition with empty space on left, 1200x630" -- this one prompt covers most thumbnails. (It took me two days to craft that single prompt line, but let's keep that between us.)

Time and Money Math

Finding an image on Unsplash and cropping it used to take 15 minutes. With AI, running the prompt three or four times and adding text in Figma takes about 10 minutes total. Cost-wise, DALL-E 3 runs about 50 KRW per image. Three or four attempts puts you at around 170 KRW. At 8 thumbnails a month, that's under 1,400 KRW.

The Failures Were More Fun

I asked it to generate a thumbnail with Korean text directly. The Korean came out as completely garbled pseudo-characters -- shapes that vaguely resemble letters. Text absolutely has to be added as a separate step in post-processing.

I also tried generating photorealistic images of someone coding. Got images with six fingers or keyboards that looked alien. Conclusion: abstract styles are the safest bet for blog thumbnails. Add a person and you're headed straight for uncanny valley.

Copyright Is Still a Gray Area

As of 2026, copyright for AI-generated images varies by country. Korea hasn't established clear guidelines yet. For personal blog thumbnails, the realistic risk of issues is low, but if you're using them commercially, definitely check each service's terms.

Better Than Nothing

For developers who can't design, AI image generation is a lifesaver. It's not perfect. But it's a hundred times better than a post with no thumbnail. The difference in click-through rates is noticeable enough -- just give it a try.