Does Substack compress images? Yes: Chrome got lossy WebP copies, a median 22 percent of the file Substack stores
On this page
Does Substack compress images? Yes. Of 159 post images from 61 publications, 158 reached Chrome 157 as a lossy WebP, a median 22 percent of the size of the stored file the image's address points to: 14 percent for PNGs, 42 for JPEGs and 7 for HEIC photos. Scored against a lossless render of the same size, those copies reached a median of 74 on SSIMULACRA 2, where 100 means no difference.
We uploaded nothing to Substack, so this is what Chrome was sent, not what the editor does with a file first.
A copy was a median 22 percent of the stored file's bytes, and never wider in pixels
| Stored as | Images | Stored file, median | Copy, median | Copy as a share of its own stored file, median (range) |
|---|---|---|---|---|
| PNG | 74 | 411 KB | 68 KB | 14% (2 to 77%) |
| JPEG | 57 | 114 KB | 54 KB | 42% (0.8 to 102%) |
| HEIC | 20 | 1,606 KB | 98 KB | 7% (2 to 69%) |
| WebP | 7 | 72 KB | 39 KB | 80% (49 to 99%) |
| All | 158 | 277 KB | 60 KB | 22% (0.8 to 102%) |
KB are 1,000 bytes. The last column is the median of each image's own ratio, so it isn't the ratio of the two medians before it. Left out: the one animated WebP, made from a GIF.
The addresses asked for 424, 848 or 1,456 pixels: 77 of the 159 for 848, 67 for 1,456 and 15 for 424. For the 83 stored files wider than their address, the copy was exactly that wide. The other 76 were no wider than their address, and their copies kept the stored width. That splits the 22 percent: the 76 lossy copies that kept their width were a median 39 percent of the stored bytes, and the 82 lossy copies cut to the address's width (the 83rd is the animated WebP) were 10 percent.
Sixty-two of the 159 stored files were over 500 KB, the largest a 5,926,350-byte JPEG. Its copy was 47,346 bytes. The copies of those 62 files had a median of 97 KB.
PNGs and HEIC photos become lossy WebP, with median scores of 74 and 72
Substack's help article on adding media lists avif, gif, jpg, jpeg, png and webp as supported files and doesn't name HEIC, yet 20 of the 159 stored files are HEIC, and Chrome got WebPs of them.
Quality scores sat close across formats: medians of 74.0 for PNG, 75.1 for JPEG, 72.1 for HEIC and 76.7 for WebP. On SSIMULACRA 2's own scale, 74 sits between "high quality" at 70 and "very high quality" at 80. Of 158 copies, 28 reached 80, 30 fell under 70 and 4 fell under 60. The lowest, 41.64, was a 1450x950 PNG of red text and a drawing on black: 103,989 bytes stored, 23,392 as a copy. Lossy WebP always subsamples colour, and our post on red text in JPG and WebP measured what that does to red text.
To check the lossless render we score against, we asked the same address for WebP at quality 100 on 25 images. The median score rose to 88.75 and none passed 93.03. A render that didn't line up with the copies would drag even these down, so the scores above come from the lossy step. Even there, the red-text graphic reached only 45.45.
The page asks for WebP, and the JPEG is only the fallback
The image addresses we saw on Substack pages carry their settings, then the address of the stored file: w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep. On our 88 loaded pages, 968 <img> tags had such an address (including post images, avatars and thumbnails); 924 sat in a <picture> whose WebP source names f_webp, and 920 had an <img> srcset that names f_auto. No page asked for f_avif. Chrome took the WebP source: all 159 post images came back as WebP, 158 lossy stills and one animated WebP made from a GIF.
The fallback doesn't follow the Accept line. We requested the <img> address of 60 post images with three Accept lines, Chrome's own, image/webp,*/* and */*. Of 180 answers, 177 were JPEG and 3 were PNG (one image, on all three lines), and none was WebP. So the page decides the format here, where other sites choose from that header.
A guide's 500 KB throttle and WebP-to-JPG claims didn't show up in the copies
One guide to Substack images, from FluxNote and updated on 23 October 2026, says "Substack throttles images over 500KB", and tells writers to avoid SVG or WebP because the CMS "still converts them to low-quality JPGs". It gives no source for either. Neither shows up in what Chrome was sent, and we uploaded nothing, so what happens at upload is untested.
Of the 62 stored files over 500 KB, one copy was over 500 KB: an animated GIF of 61 frames, 2,391,589 bytes stored and 1,010,392 as an animated WebP. Seven stored files were WebP, Chrome got a WebP for all seven, and the copies were a median 80 percent of the stored bytes. JPEG appeared only at the fallback address. Substack's help centre states no size limit for post images in any of its 396 articles (read on 16 November 2026), and the media article lists WebP and AVIF among the supported files.
The stored file is public and can name a camera
The stored file behind each of the 159 post images is public: we fetched all 159 without logging in. Wix works the same way. Sixty-four of the 159 held EXIF and seven named a maker; none held GPS coordinates. All 159 copies carried an EXIF block with no maker and no GPS; the two we read listed the orientation, 72 dpi and the pixel size.
What to upload
- Exporting wider than 1,456 pixels buys nothing on the page: none of the 159 post-image copies Chrome loaded was wider, and none of the
<img>tags withc_limitaddresses on the 88 pages offered a file wider than 1,456 pixels in its srcset, so a 2x screen would have no wider one to pick. We didn't open the expanded view of an image. - Upload the best file you have. Substack makes the copy Chrome gets itself, at a median 22 percent of the stored bytes, so the stored file's size isn't the size Chrome downloads. A lossy pass of your own in front of Substack's is a second lossy pass, and on Next.js the unedited photo at a lower quality setting scored higher than the same photo compressed first in 40 of 40 comparisons at no more bytes (our test of compressing before a CDN re-encodes). We uploaded nothing, so that is untested on Substack.
- To check your own post, copy an image address from it in Chrome (right-click, Copy image address) and run this in a terminal with curl and a POSIX shell (we ran it under sh, bash and zsh on Linux):
URL='paste the image address here'
curl -fsS -o copy.webp "$URL"
curl -fsS -o stored "$(printf '%s' "$URL" | sed -e 's#.*/https%3A#https:#' -e 's#%2F#/#g')"
wc -c copy.webp stored
The first size is the copy and the second is the stored file, whose address the sed line cuts out of the one you pasted.
How we measured
Fifteen of Substack's own sitemap files (sitemap-tt-1 to -15) listed 243,620 publication home pages when we read them on 16 November 2026, and 124,251 had been updated since 1 October 2026. Nine more files list about 134,000 others, which the draw left out.
We drew 900 at random with a seeded draw. Of those, 622 gave us a post list from their own sitemap (270 answered 429, eight had no recent post), and we took one post updated since 1 September 2026 from each. We drew 360 of the 622 for the browser and loaded the pages in Chrome 157.0.8103.78 at 1440x900 with a device pixel ratio of 1, each scrolled to the end. The first 46 ran in alphabetical order; after a restart the rest ran in a seeded random order. Custom domains, phones, 2x screens, Notes, email, the app and other browsers are not in the sample.
The run ended at 99 of the 360 pages. Substack still answered 429 to 11 of the 99 when it ended, and we didn't go around it. That left 88 loaded pages, and 61 of those held at least one post image: an address with w_, c_limit and f_webp whose stored file sits in Substack's public post-media storage. Avatars, thumbnails and the header banner are left out.
For each image Chrome requested, we fetched the same address with Chrome's Accept line, the stored file named in it, and the same address with its settings replaced by f_png, a lossless render of the same resize. Our downloads had the same byte counts as the Content-Length Chrome saw, for all 159. We decoded the copies with dwebp (libwebp 1.5.0) and scored them against that render with ssimulacra2 from libjxl 0.12.0, so the score measures the lossy step and not the resize. All 665 addresses of files in that storage (post images, avatars and thumbnails) returned identical bytes when we asked again 1 to 24 minutes later.