We ran 54 business photos through Picture Optimizer's own pipeline and measured every file. As AVIF at our High preset, resized to 2,048 px wide with metadata stripped, the set went from 183.0 MB to 16.2 MB, a 91.1% reduction. The lowest quality score (SSIM against the resized original) was 0.9908. This post shows the method, the totals, the results by type of business and the photos that did not shrink much.
Key takeaways
- Total: 183.0 MB of originals became 27.1 MB as JPEG, 26.7 MB as WebP and 16.2 MB as AVIF.
- The median photo shrank by 93.8% as AVIF; the range was 24.1% to 99.0%.
- Every AVIF scored SSIM 0.9908 or higher; 41 of the 54 scored 0.995 or higher.
- Resizing did a lot of the work: output had about 15% of the original pixels.
- Texture and starting size, not business type, explained the outliers.
What we tested
The photos. 54 free-license photos from Unsplash, picked to look like what small businesses actually publish: a plumber under a sink, a dental treatment room, a bakery display, a law firm signing, a boutique interior, a food blogger at work and so on. They are not customer photos. The groups were local services (20), trades (10), healthcare (7), online stores (5), professional services (4), blogs (3), agencies (2), and one each for a photographer, an SEO team and a developer.
The originals. All were JPEGs. The median was 24 megapixels and about 3.0 MB; the smallest was 0.75 MB and the largest 10.1 MB. Every original was wider than 2,048 px (from 2,202 px to 11,648 px).
The pipeline. Each photo went through the same code the app uses, with one setting for all: the High preset (JPEG and WebP at quality 82, AVIF at quality 55 with 4:4:4 chroma), resized to 2,048 px wide, metadata stripped. We saved JPEG, WebP and AVIF versions of each, plus AVIF copies at 300, 768 and 1,024 px for responsive use. The run finished on 17 October 2026.
The quality check. SSIM, the structural similarity index from Wang, Bovik, Sheikh and Simoncelli (2004), scores two images from 0 to 1, where 1 means identical. We compared each AVIF with the original resized to the same dimensions, at 1,024 px. That isolates what the encoding changed; it does not score the resize itself, which is intentional.
Units. MB means 1,048,576 bytes, the way our tool reports file sizes. Percentages are calculated from bytes, not from rounded MB.
The totals
Version | Total bytes | Total | Smaller than originals by |
|---|---|---|---|
Originals (JPEG) | 191,915,315 | 183.0 MB | n/a |
JPEG, High, 2,048 px | 28,383,973 | 27.1 MB | 85.2% |
WebP, High, 2,048 px | 27,945,242 | 26.7 MB | 85.4% |
AVIF, High, 2,048 px | 17,003,431 | 16.2 MB | 91.1% |
Method for the last column: 1 − (version total ÷ 191,915,315). For AVIF that is 1 − 17,003,431 ÷ 191,915,315 = 0.911.
A total can be dominated by a few big files, so we also looked at single photos. The median AVIF was 220 KB, and the median photo shrank by 93.8%. Eleven AVIFs were under 100 KB; seven were over 500 KB.
Results by type of business
Group | Photos | Originals | AVIF | Reduction | Lowest SSIM |
|---|---|---|---|---|---|
Local services | 20 | 67.7 MB | 6.8 MB | 90.0% | 0.9928 |
Trades | 10 | 39.2 MB | 3.0 MB | 92.3% | 0.9925 |
Healthcare | 7 | 17.4 MB | 3.0 MB | 83.0% | 0.9913 |
Online stores | 5 | 15.9 MB | 0.9 MB | 94.4% | 0.9909 |
Professional services | 4 | 11.8 MB | 1.0 MB | 91.8% | 0.9908 |
Agencies | 2 | 13.5 MB | 0.3 MB | 97.5% | 0.9937 |
Blogs | 3 | 11.1 MB | 0.8 MB | 93.1% | 0.9955 |
Photographer, SEO team, developer | 3 | 6.6 MB | 0.5 MB | 91.8% | 0.9972 |
Healthcare looks weaker, but one photo explains it. Without the yoga studio photo, the other six healthcare photos went from 15.0 MB to 1.1 MB, a 92.6% reduction. Business type doesn't predict compression; image content and starting size do.
Where the savings came from
Two things happened to every photo: it got fewer pixels, and the pixels it kept were encoded more efficiently.
Fewer pixels. The 54 originals held 1,262 megapixels in total. At 2,048 px wide they held 194 megapixels, about 15% of the original. The biggest single reduction in the run, 99.0%, was an auto mechanic photo that started at 11,648 × 8,736 px and 7.4 MB and ended at 79 KB.
Better encoding. Per pixel, the median original used about 1.23 bits; the median AVIF used about 0.52 bits. That is the format and quality setting at work.
We can't fully separate the two effects from this data, because we didn't save a resized-only copy at the original's quality. What the data does show is that format still mattered after resizing: at the same size and preset, the AVIF totals were 40.1% smaller than JPEG and 39.2% smaller than WebP.
The photos that barely shrank
The outliers are the most useful part of a test like this.
- Yoga studio, 24.1% smaller (2.46 MB to 1.86 MB). The original was only 2,202 px wide, so resizing removed little, and the backlit sunset scene has a film-style grain across the whole frame. Grain is random detail, which encoders cannot predict well. SSIM was 0.9913.
- Food photo for a blog, 74.7% smaller. At 3,136 px wide it was one of the narrower originals, so resizing removed less.
- Landscaping, 75.6% smaller. A portrait photo of a lawn: grass is close to noise for an encoder. Its AVIF was 1.5 MB, the second largest in the run.
- Tattoo studio, 78.2% smaller. A dark, grainy close-up of detailed line work.
The pattern: if a photo is already near your target width, or is full of fine random detail (grass, gravel, grain, foliage), expect a smaller percentage. Don't drop the quality setting to force a number; check those files by eye instead.
What we would tell a site owner
- Resize first. If your theme shows images at 1,200 px, a 6,000 px upload is mostly wasted bytes.
- Prefer AVIF for photos, with a fallback. It was the smallest format for every photo here. Caniuse reports about 96% global support; MDN recommends a fallback for the rest.
- Don't assume WebP beats JPEG. At these settings WebP was larger than JPEG for 20 of the 54 photos, and only 1.5% smaller in total.
- Measure, then trust the setting. Once a preset holds up on your hardest photos, use it for the batch.
Limits of this test
One preset, one target width, one encoder version, and stock photos rather than phone snapshots. Phone photos are often noisier and may compress less. SSIM was measured for AVIF only, not for the JPEG and WebP versions. Treat the numbers as a well-documented example, not a promise for every library.
Run the same test on your photos
Our bulk image compression tool uses the same pipeline and presets. The Free plan covers your first 15 photos, once, as AVIF or JPG, and resizes photos over 3 MB to fit. Starter ($19 a month) keeps originals up to 5 MB and adds WebP output and AI alt text; Pro keeps originals up to 30 MB. More on formats: convert JPG to AVIF and compress AVIF files.
Keep reading
Sources
Questions
Short answers to what people ask about this.