Progressive JPEG Saved Us 3 Percent. Full Color Chroma Cost 24 Percent More For About 1 dB

Two JPEG settings everyone argues about, measured on the same 400 pictures. One is a small free win. The other is a quarter more bytes.

Published September 19, 2026 · RealAIGirls · About a 5 minute read

There are two JPEG settings people love to argue about. Progressive encoding, which is supposed to make files smaller for free. And chroma subsampling, the setting that decides how much color detail gets thrown away.

We stopped arguing and measured both. On September 19, 2026 we took a random sample of 400 images from the realaigirls.com image folder, scaled each one to fit inside 1200 by 1200 pixels, and saved every one of them four ways at JPEG quality 85.

The short version: progressive is a small win and you should take it. Full color chroma is a big cost for a change most people can’t see.

The Four Saves

Same 400 pictures, same quality number, four different settings. Totals for all 400:

Baseline, 4:2:0 chroma, no optimize: 43.34 MB. This is what a lot of tools do by default.
Baseline with optimized Huffman tables: 42.51 MB. That’s 1.9 percent smaller.
Progressive with optimized tables: 42.03 MB. Another 1.1 percent, and 3.0 percent smaller than where we started.
Progressive with 4:4:4 chroma: 52.05 MB. That’s 23.9 percent bigger than the progressive 4:2:0 file.

Progressive Is Real, But It’s Three Percent

You’ll see claims that progressive JPEGs are 10 percent smaller or more. On our images it didn’t come close. The median image got 2.6 percent smaller going from plain baseline to optimized progressive.

It also isn’t a sure thing on every file. Progressive beat optimized baseline on 301 of the 400 images, and lost on the other 99. The bigger the file, the more reliable it got:

Files under 60 KB: progressive won on 22 of 24.
Files from 60 to 120 KB: it won on 172 of 268, the shakiest group.
Files over 120 KB: it won on 107 of 108, and saved 3.6 percent overall.

Three percent won’t change your life. But it costs nothing, it loads as a blurry full picture instead of a top down reveal, and across a gallery of a few thousand images it adds up. Turn it on.

Full Color Chroma Is A 24 Percent Tax

JPEG usually stores color at half resolution in both directions. That’s 4:2:0. Your eyes are much sharper on brightness than on color, so most of the time you never notice. Setting it to 4:4:4 keeps every bit of color detail.

Here’s what that bought us. We measured PSNR against the resized original for each file. The median was 41.94 dB at 4:2:0 and 43.11 dB at 4:4:4, a median gain of about 1 dB. The median file cost 24.3 percent more bytes for it.

And that cost barely moved with file size. Small files paid 25.1 percent, mid size files 24.5 percent, big files 22.8 percent. You pay about a quarter more no matter what.

Is it ever worth it? Sometimes. Saturated red edges, thin colored text and neon lighting are where 4:2:0 smears first. If your image is a portrait with skin, hair and a soft background, which is most AI images on this site, you’re paying 24 percent for detail that mostly isn’t there to lose.

What We’d Actually Set

Always optimize the Huffman tables. 1.9 percent smaller, same pixels, no downside.
Save progressive. Another 1.1 percent on average, and near certain on bigger files.
Leave chroma at 4:2:0 unless the picture is mostly hard colored edges or text.
If bytes really matter, change formats instead. Tweaking JPEG settings got us 3 percent. Our WebP versus JPEG measurements found far bigger savings from switching.

Method: 400 images drawn at random (seed 19) from the realaigirls.com image folder, decoded to RGB and fit inside 1200 by 1200 pixels with Lanczos resampling. Each was saved with Pillow 12.3.0 at JPEG quality 85 four ways: baseline 4:2:0; baseline 4:2:0 with optimize; progressive 4:2:0 with optimize; progressive 4:4:4 with optimize. PSNR was computed over RGB against the resized image. Run September 19, 2026.