Does Framer compress images? Yes: a WebP first, then an AVIF about 30 percent smaller
On this page
Yes. On 15 November 2026 Chrome 157 loaded 300 sites on Framer's free framer.website domain, and 288 of them served raster images: 4,638 responses in all, of which 2,380 arrived as AVIF, 2,131 as WebP and 127 as a JPEG, PNG or GIF. Where we saw both an AVIF and the WebP for one address, the AVIF weighed a median 0.695 of the WebP on addresses we made fresh and 0.709 on ones we replayed.
Framer's own pages give three answers. Its glossary says "Framer serves WebP images to supported browsers automatically". The help page, updated 15 September 2026, says most images become AVIF at quality 80, and the engineering blog, from 30 May 2024, says "All images on Framer are now served as AVIF", then describes a WebP that goes first. We ran the help page's table against the files. If you only want the advice, skip to "What to do".
Chrome got a first-request WebP for 42 percent of its raster responses
Chrome 157's Accept header lists JPEG XL first (the line our earlier post printed, from Chrome 153, did not), and none of the 5,605 image responses used it. Setting aside 967 SVG responses, the other 4,638 were:
| What Chrome got | Responses | Share |
|---|---|---|
| AVIF, cacheable for a year | 2,380 | 51.3% |
WebP with Cache-Control: max-age=0 |
1,968 | 42.4% |
| WebP cacheable for a year | 163 | 3.5% |
| JPEG, PNG or GIF | 127 | 2.7% |
Framer's blog says it answers an address's first request with a WebP carrying max-age=0, so the CDN asks again, and gives the second request an AVIF. We call any response with that header the first-request kind: 2,051 of the 4,638 (1,968 WebP, 51 JPEG, 32 PNG). On 229 of the 288 sites at least one response was that kind, on 85 every raster response was, and on 59 none. We can't tell whether our crawl was the first request for an address or the CDN had dropped a copy it held.
AVIF follows, about 30 percent smaller, but not on a schedule
To watch the switch, we made 60 new addresses (all scale-down-to=1024 copies) by giving existing images a new width and height, which the CDN counts as a different address and which change nothing in the file. We asked for each with Chrome's Accept header at 0, 30 and 60 seconds, then at 90 and on to 600. At 0, 30 and 60 seconds, 57 of the 60 returned a first-request WebP (two returned a JPEG and one a PNG). At 90 seconds 56 of those 57 returned AVIF, and they still did at 600.
Two larger runs were messier. We replayed 1,952 distinct raster addresses from the crawl, asking three times a few minutes apart, and AVIF came back for 603, then 953, then 1,352 of them. For 164 and then 225 addresses an AVIF was followed by a WebP at the next pass.
On 120 other fresh addresses, asked again at once and then four more times, AVIF first came back at 58 seconds: 22 of 120 requests made 26 to 90 seconds after the address's first response returned it, and 109 of the 119 recorded, made 75 to 149 seconds after, did. At about 7 minutes 61 returned AVIF and 55 the first WebP again, byte for byte (the other 4 were PNG or GIF). Between 15 and 22 minutes 113 of 120 returned AVIF. We don't know why some addresses went back to the first WebP here and none did in the 60-address run.
Where an address returned a WebP first and an AVIF later, the AVIF weighed a median 0.695 of the WebP on 111 fresh addresses (middle half 0.658 to 0.739, smaller on 104) and 0.709 on 839 replayed ones. Framer's blog says ~20% smaller, and our Wix post found Wix's AVIF a median 16 percent smaller than its WebP on JPG images. The help page gives 80 for AVIF and 90 for the first WebP, so the comparison is of bytes at two different qualities; our AVIF-against-WebP post holds the score fixed. SSIMULACRA 2 scored the AVIF a median 80.1 against the WebP (109 of 111 pairs, 4 under 70).
Framer's help page, claim by claim
| Framer's help says | We ran | What came back |
|---|---|---|
| Copies at 512, 1,024, 2,048 and 4,096 px by the larger side, never upscaled | scale-down-to at each size on 40 framer.com images |
158 of 159 readable results matched the smaller of that size and the source. A 1,062 px source asked for 1,024 came back at 1,062 |
| A browser without AVIF gets WebP | Chrome's line without AVIF and JPEG XL | The first-request WebP's bytes for 114 of the 114 addresses that first returned a WebP |
| One with neither gets "a recompressed version of the original" | Accept: */* |
JPEG at libjpeg quality 90, 4:2:0, progressive, no EXIF for 39 of the 41 JPEG files returned |
| An optimized file larger than the original: the original is served | A 7,952 px JPEG at quality 73 | The same 4,118,848 bytes with Chrome's line at two addresses and with Accept: */*; its 4,096 and 2,048 px copies were smaller WebP files. None of the 33 JPEG files the replays returned had the quality-90 tables |
| Auto (Lossless): lossless WebP | 66 addresses with lossless=1, which we take to be that setting |
60 were lossless WebP. Six JPEG addresses each returned a quality-40 JPEG at the source's own 1,280 px, though they asked for scale-down-to=512 |
| Animated images stay lossy WebP | 6 GIF addresses | 4 animated WebP with lossy frames, 2 GIFs |
The page held in the first four rows, apart from a few addresses, and the last two show exceptions. We uploaded nothing to Framer, so quality 80 and 90 are Framer's figures, not ours, and the fourth row fits the help page's rule without showing that size was the reason.
Images without a srcset get the full-size source
scale-down-to resizes. A srcset lists the copies narrower than the source (512, 1,024, 2,048 or 4,096 px), then the source at full size at the plain address, the one with no scale-down-to. width=800, height=400 and width=800&height=534 on a 7,900 px photo returned 7,900 px, and in 3,188 of 3,188 plain entries the width in the address equals the entry's own width, the source's.
Of 4,311 (site, image) pairs with a readable source width, 275 are wider than 4,096 px, on 78 of 299 sites. 237 of the 275 have a srcset. For the 181 of the 237 that both the pixel-ratio-1 and pixel-ratio-2 loads saw, on 63 sites, Chrome downloaded no file wider than 4,096 px (pixel ratio 2 is two device pixels per CSS pixel, in the same 1440 px window). An image without a srcset has only the plain address, and Chrome loaded 11 such images, of the 3,930 it loaded at pixel ratio 1, at 4,160 to 8,000 px.
What to do
- Skip converting to WebP or AVIF first. Framer's page calls that usually unnecessary, and JPEG and PNG sources reached AVIF without it: 43 of 43 JPEG and 60 of 63 PNG addresses within 22 minutes.
- Read
Cache-Controlin DevTools. A WebP withmax-age=0is the first answer; on our fresh addresses the first AVIF we saw came 58 seconds to 7 minutes after the first WebP, and on some it hadn't come when we stopped. A page scan lists the images a page loads, with the sizes in eachsrcset. - Resize big photos before you upload them. Framer's page says to upload unmodified, but an image with no
srcsetis served at the source's full size: Chrome loaded 11 such images at 4,160 to 8,000 px. Iminify resizes by pixels. - Use Auto (Lossless) only if compression shows. Framer's resolution page says to, and that it can make the file 5 to 10 times bigger; on a JPEG address it returned a full-size JPEG even for a 512 px copy.
How we measured
On 15 November 2026 we took the 2,565 hosts under framer.website in Common Crawl's crawl of 4 to 17 September 2026, kept the 2,221 whose home page answered 200, was made by Framer and named an image on framerusercontent.com, dropped 70 whose title or name matched a gambling or adult word, and drew 300 with a fixed seed. About one in five of the 2,151 left says template, preview or demo in its title or name. Chrome 157.0.8103.78 loaded each home page once at 1440 by 900 pixels with the cache off and scrolled down the page; 299 answered 200.
We asked for addresses with Chrome's own Accept header, with Chrome's line minus AVIF and JPEG XL, or with */* alone. Formats come from Content-Type in the crawl and from the first bytes in our own requests, JPEG quality from the quantisation tables, and the scores from ssimulacra2 in a pinned encoder image. The 1,952 replayed addresses are the distinct raster ones behind the 2,051 first-request responses. The 74 sites whose srcset reached past 4,096 px were loaded again at pixel ratio 2.