I Tried Getting AI to Make My Thumbnails...

> TLDR

$ trained a custom AI model on 113 photos of myself

$ also tried Google's image AI with just a reference photo

$ both were... ok. great for sketching ideas. not for real thumbnails.

I had this idea. What if I could train an AI model on my face so it could generate YouTube thumbnails for me? No more photoshoots, no more editing. Just type a prompt and get a thumbnail that looks like me. Or at the very least, a fast way to sketch out ideas before I shoot.

So I tried two different approaches. Here's what I found...

Approach 1: Fine-Tuning an Image Model

I uploaded 113 photos of myself to Replicate and trained a LoRA model. That's basically teaching an AI what I look like so it can generate new images of me. It took hours to edit the photos, add descriptions, and train.

REPLICATE LORA TRAINING

what: fine-tune an image model on your face

photos: 113 uploaded

time: hours of training

cost: ~$2 to train

The results were... ok. It got the vibe right. The lighting, the background, my beanie. My face looked funky in a couple of them, but it was pretty close. The hands? Yeah those ain't it (but I doubt anyone other than me would ever notice). The guitar looked good, but it wasn't my guitar. I definitely gave it a lot of photos of my guitar. And that last picture on the bottom right — what even is that? A 9-string guitar? Oof.

LORA RESULTS — TRAINED ON 113 PHOTOS

VERDICT: MEH

To be fair, my thumbnails are a little more complex than just my face. I have shots with my guitars, close-ups of my hands on the guitar, etc. I uploaded a ton of photo examples with the guitar and hands too, but yeah it just wasn't quite there.

Approach 2: Just Give AI a Reference Photo

Then I tried something completely different. Google's image generation model (Nano Banana). Instead of training a custom model, you just give it a couple of photos as a reference and tell it what you want. No training. No uploading 100+ photos. No waiting hours. Just "here's what I look like, now make this."

GOOGLE GEMINI (NANO BANANA)

what: generate images from a reference photo

photos: just 1-2 needed

time: ~15 seconds per image

cost: pennies

The results were... also ok. I was surprised at how much it nailed the guitar. I mean, that is spot on. I gave the LoRA model a ton of guitar photos and it still didn't come close. The expressions are a little funky, but the face and hands are pretty good. It's close - but would I use any of these as an actual thumbnail? Not a chance.

NANO BANANA RESULTS — FROM 2 REFERENCE PHOTOS

VERDICT: ALSO MEH

So... Which One Won?

Neither. That's the honest answer.

LORA FINE-TUNE

hours + ~$2

113 photos

face was close-ish

hands were rough

guitar was wrong

good for sketching ideas

NANO BANANA

fast / pennies

1-2 photos

face was close-ish

guitar was spot on

still has AI look

also good for sketching ideas

Both tools did the same thing — gave me something that kinda looks like a thumbnail. But... I would never use it. And honestly? That's fine. Because that's not really the point.

Here's What I Actually Use Them For

// real talk

Before I shoot a thumbnail, I usually sketch ideas on a notepad. What pose, what angle, etc. Now I can type a prompt instead of drawing stick figures. I get a rough idea in 15 seconds that helps me plan my actual photoshoot.

That's it. That's the use case. AI sketches, not AI replacements.

As a quick brainstorming tool, it's awesome. Just a quick "would this pose look good?" before I set up all my crap? It's really useful for that.

I still shoot my own thumbnails. I still have my editor clean them up. AI hasn't replaced that process. But it does allow me to work significantly faster than I did before. And saving time is a win.

<< BACK TO POSTS