JPG vs WebP vs AVIF: We Tested 33 Product Photos (2026)

JPG vs WebP vs AVIF, in short: on 33 of our own product photos (about 1000 px), WebP at quality 80 came out at a median 44% of a quality-85 JPG’s size with near-identical SSIM (0.9928 vs 0.9935). AVIF at quality 50 was smaller still (22%) but scored lower (0.9843). PNG files were 5.9 times the JPG’s size.

More on this topic: Photo editing how-tos, every guide in one place.

Most format advice repeats the same vendor figures. We wanted numbers from images a seller would actually upload, so we took every distinct product and scene photo on this site, encoded each one eleven ways, and measured file size and similarity to the original. This page gives the method, the full results with medians and spread, what we could not test, and what it means for Amazon, Etsy, Shopify and your own store. The per-image results are at the end of the post, so the numbers can be checked.

Is WebP smaller than JPG for product photos?

Yes, for almost every image in our test. WebP at quality 90 was smaller than JPEG at quality 85 for 32 of 33 images (median 62% of the size) while scoring a higher median SSIM (0.9954 vs 0.9935). WebP at quality 80 was smaller for all 33, at a median 44% (middle half of images: 40–55%).

The one image where WebP q90 was not smaller (gen-wallet.webp, a busy flat-lay on linen) came out at 368.7 KB against 295.3 KB for the JPG, so test your most detailed shots rather than assuming WebP always wins.

Is AVIF better than WebP for product images?

AVIF made the smallest files, but its lowest settings also had the lowest SSIM. AVIF q70 averaged 39% of the JPEG q85 size with SSIM 0.9913; WebP q80 averaged 44% with SSIM 0.9928. So at q70 AVIF was a little smaller but also lower-scoring than WebP q80 on these images. AVIF q50 (22%, SSIM 0.9843) and q35 (14%, SSIM 0.9741) are where AVIF pulls far ahead on size, and where its SSIM falls furthest: the worst image at q35 scored 0.7944.

For AVIF to match JPEG q85 by SSIM you would need a setting above q70 for most of our images: only 4 of 33 reached it at q70. We did not test higher AVIF settings, so we can’t say what size AVIF needs to match JPG q85. And for uploads it matters less than it sounds: of Amazon, Etsy and Shopify, none lists AVIF as an accepted upload format. If you have AVIF files, our free AVIF to JPG converter turns them back into JPGs.

What JPG quality should you use for product photos?

Quality 85 was the sensible middle in our data. Going from q85 up to q95 made files 176% as large (median 111.0 KB vs 63.9 KB) for a small SSIM gain. Going down to q75 made them 76% as large (median 47.6 KB) with SSIM 0.9904. At q60 the median SSIM fell to 0.9830 and the worst image dropped to 0.8554.

How much bigger is PNG than JPG for photos?

About 5.9 times bigger at the median. Optimized PNG files were a median 359.4 KB (middle half 170.4–783.3 KB), against 63.9 KB for JPEG q85. PNG is lossless, so its SSIM is exactly 1, against 0.9935 for the q85 JPG: a small gap for a much heavier file. Keep PNG for transparent cutouts, logos and text graphics.

Results: file size and SSIM for every setting

Median and interquartile range (the middle 50% of images) across all 33 images. “vs JPEG q85” is the median of each image’s size divided by its own JPEG q85 size. SSIM is 1.0000 for an identical image; higher is closer to the original.

How many images fit common size limits?

Out of 33 images at their original size of about 1000 px. Larger images need more bytes, so treat these as best-case for 1000 px uploads, not for 2000 px marketplace images.

Does stripping metadata or embedding a color profile change file size?

Not for our files. None of the 33 sources carried EXIF, XMP or an ICC profile (0 bytes of metadata in total), so stripping saved nothing. Embedding a standard sRGB ICC profile in a q85 JPG added 606 bytes, a median 0.95% of the file. Phone photos are different: they usually carry EXIF and location data, which our free metadata remover strips. Saving the JPG as progressive instead of baseline came to a median 98.4% of the baseline size.

How we tested

• Sample: 33 RGB product and scene images from this website’s own pages (apparel, food, jewelry, a serum bottle, mugs, a bakery, a sneaker, a wallet), long side 800–1024 px. We left out logos, icons, social share images, tool screenshots, documents, images with text overlays and any image that was a copy or tool-processed variant of another.
• Reference: each source decoded to 8-bit RGB. Every encode was decoded again and compared with that reference.
• Metric: SSIM on the luma (brightness) channel, BT.601 weights, 11×11 Gaussian window with sigma 1.5, the standard constants from the original SSIM paper.
• Encoders: Pillow 11.3.0 with libjpeg-turbo, libwebp 1.5.0 and libavif 1.3.0. JPEG with optimized Huffman tables and 4:2:0 chroma subsampling (Pillow’s default); WebP at method 4; AVIF at speed 6; PNG with optimize on. Quality numbers are each encoder’s own scale, which are not equivalent between formats.
• Settings: JPEG q95, JPEG q85, JPEG q75, JPEG q60, WebP q90, WebP q80, WebP q70, AVIF q70, AVIF q50, AVIF q35, PNG optimize.
• Run: 2026-10-26, offline, on one Mac, with a short Python script using Pillow and NumPy. The settings above are everything needed to repeat it. Download the script and the raw results (JSON).

What are the limits of this test?

• Already-compressed sources. Every input was stored on our site as a lossy WebP. Re-encoding a decoded WebP may flatter WebP, and none of the sources were camera originals.
• Small images. All sources are around 1000 px. None was 2000 px or larger, so we skipped the planned 1000 vs 2000 px comparison and can’t give file sizes for 2000×2000 marketplace images.
• A narrow mix. 14 of the 33 images show the same serum bottle on different backgrounds, and many are clean studio shots or AI-made scenes from our own tools. Noisy low-light phone photos would compress differently.
• One metric. SSIM on brightness only. It doesn’t measure color shifts, and we ran no viewing test with people.
• Default encoder speeds. Slower WebP and AVIF settings, or other encoders, can make smaller files. We didn’t time the encodes.

What does this mean for Amazon, Etsy and Shopify uploads?

For marketplace uploads, the format question is mostly settled by what the platform accepts, not by our numbers.

For the full format guide, including transparency and color profiles, see JPG vs PNG vs WebP for product images. For hitting a size cap, see how to reduce image file size and website image sizes.

Frequently asked questions

Is WebP better than JPG for product photos?

For your own website, yes: in our test of 33 product photos, WebP at quality 80 was a median 44% of the size of a quality-85 JPG with almost the same SSIM. For marketplace uploads, use JPG: Amazon doesn’t list WebP as an accepted format and Etsy’s support isn’t confirmed.

Is AVIF smaller than WebP?

At low quality settings, yes. AVIF q50 was a median 22% of the JPG size against 44% for WebP q80, but its SSIM was lower (0.9843 vs 0.9928). At AVIF q70 the files were 39%, close to WebP q80.

What JPG quality is best for Amazon and Etsy?

Quality 85 is a good default. In our test q95 files were 176% the size of q85 files, and q75 files were 76%, with SSIM falling from 0.9935 to 0.9904.

Should product photos be saved as PNG?

Only if they need transparency or contain sharp text. Our optimized PNG photos were a median 593% of the size of the same image as a q85 JPG.

Does removing metadata make images smaller?

Only if the image has metadata. Our 33 web images had none, so stripping saved nothing; phone photos usually carry EXIF data that is worth removing for privacy.

Can I check these results?

Yes. The per-image results are in the table at the end of this post, and the method section lists every setting, so the test can be repeated offline with Pillow and NumPy.

Per-image results

File size in KB, with SSIM in brackets, for four of the eleven settings. The full set is in the summary table above.