What size should Ghost images be? Ghost shrinks...

What size should Ghost images be? Ghost shrinks uploads past 2,000 px wide, re-saves JPEGs at quality 80 and keeps your original

By SJ
8 min read
Diagram: a 13.1 MB photo.jpg of 8160 x 6144 pixels goes into Ghost, which shows a 673 KB copy of 2000 x 1506 pixels and keeps the 13.1 MB original as photo_o.jpg.
On this page

We uploaded a 13.1 MB photo from a Pixel 8 Pro to a Ghost 6.77.0 test site. Ghost kept a 673 KB JPEG, 2,000 pixels wide and without location data, as the copy it shows. The 13.1 MB file stayed in the same folder, location included, at an address that opens without logging in.

So upload at 2,000 pixels wide or more. Wider files get shrunk, JPEGs get re-saved at quality 80 unless that would make them bigger, and your original stays beside the copy, location included. Strip the location before you upload if that matters.

Ghost shrinks a wide upload to 2,000 pixels and saves JPEGs at quality 80

Ghost's configuration docs, read on 17 November 2026, list the steps: resize to "2000px max width", "JPEGs are compressed to 80% quality", "Metadata is removed", and "The original image is kept with the suffix _o." We checked them on the photographs from our JPG-size post:

Photo Upload Ghost's copy Score
Café, Pixel 8 Pro 8160x6144, 13.1 MB 2000x1506, 673 KB 71.10
Cemetery, Galaxy S24 Ultra 8160x5100, 20.3 MB 2000x1250, 495 KB 72.85
Palenque, Fujifilm X-T4, saved on its side 6240x4160, 16.6 MB 2000x3000, 1,226 KB 72.20
Crocuses, Pixel 8 4080x3072, 6.5 MB 2000x1506, 877 KB 76.63
Station, Sony A7 III 6000x4000, 14.9 MB 2000x1333, 322 KB 72.50
Milky Way, Canon EOS R6 3648x5472, 27.3 MB 2000x3000, 1,171 KB 63.80
Aurora, Nikon Z 6 5978x3977, 23.6 MB 2000x1331, 295 KB 53.43

The score is SSIMULACRA 2, where higher means closer to the upload and 80 means differences you can't see side by side at full size. Five of the seven scored 71 or more. The two night skies did not.

The quantisation tables of all 51 JPEG copies Ghost re-saved match mozjpeg 4.1.5 at quality 80. None kept a camera name, a colour profile or an orientation tag, and none of the 20 re-saved JPEG files that carried GPS coordinates kept them. Ghost turned the Fujifilm photo upright and converted the Pixel 8's Display P3 photo to sRGB before dropping the profile, so the colours held. Six repeated uploads gave byte-identical copies.

Twenty-one of the 22 files wider than 2,000 pixels came back 2,000 wide. The exception is in the section on files Ghost keeps.

Most JPEGs under 2,000 pixels wide get re-saved too

A forum user's summary of emails from Ghost(Pro) support, posted in November 2024, says images are "used as is if they are 2000px in width or smaller", so you should "optimize them on your end." That isn't what 34 of our 40 JPEGs at or under 2,000 pixels wide got. Our 1,800-pixel Fjord photo went from 852,461 to 369,917 bytes. The 34 came back smaller, among them a copy saved at quality 1 and files mozjpeg and jpegli had already compressed; the other six are in the section on files Ghost keeps. In the Ghost(Pro) sample, the stored copy was smaller than the original for all 37 JPEG originals at or under 2,000 pixels wide.

We sent three photographs four ways and scored Ghost's copy of each:

Upload Café Crocuses Palenque
The original 71.10 76.63 72.20
Resized to 2,000 px at quality 95 69.44 76.17 70.97
Iminify Smart at 2,000 px 67.76 73.99 68.46
Iminify Lossless, copy compared with the original's identical identical 23 bytes apart

The bytes of Ghost's copies stayed within 4.1 percent of the original's. The original scored highest, ahead of the resized and Smart files by 0.5 to 3.7 points, so send Ghost the file you have. Each arm also used a different resizer, so we can't say which step costs what.

Iminify is ours. Its Lossless level repacks the JPEG with jpegtran, which changed no pixel in our check of 25 files. Ghost's copy matched the original's on two photographs. On Palenque, which Lossless turned upright, it differed by 23 bytes.

Your original stays in the folder, and so does its location

Ghost saves your upload beside the copy with _o before the extension, so photo.jpg has a photo_o.jpg. For all 70 files the _o copy was byte-identical to what we sent, and each one opened at its address without logging in. Ghost's configuration docs name the file and don't say it's public. Wix has a public file behind every image too.

On the 60 Ghost(Pro) sites, 140 of 165 originals opened, at 52 sites. The first 64 KB of each showed GPS coordinates in at least 9 photographs at 6 sites and a camera make or model in at least 16 at 13 sites.

To keep a location out of the public file, delete it from a copy of the photo before you upload. This ExifTool 13.25 command removes GPS coordinates from EXIF and XMP and left every decoded pixel unchanged; on three photographs it changed the files by under 250 bytes and left no GPS tag in any group:

exiftool -gps:all= "-xmp:gps*=" -overwrite_original photo.jpg

It doesn't remove a place name an editor typed into the file. It keeps the orientation tag and the colour profile, so Ghost still turns and converts the photo: its copies of the stripped Café and Crocuses photos and of a stripped Orientation 6 test photo matched the ones from the unstripped files. Iminify's Lossless level also drops the location by default, and those uploads left no _o with GPS.

A PNG, WebP or JPEG stays as you sent it when Ghost's re-save would be bigger

Ghost's image code returns the re-saved image only if it's smaller than the upload. Four of our seven PNGs failed that test, among them a 3.5 MB photograph. So did a 2,400-pixel animated WebP, the one wide upload that kept its size, and so did six of our 57 JPEGs.

We made those six with sharp, the library Ghost uses, from the Fjord photo at 1,200 pixels wide and quality 20 to 70, and gave them the Pixel photo's GPS tags. They weighed 62 to 168 KB. The two at quality 75 and 80 (186 and 214 KB) were re-saved to 175 and 181 KB.

A file Ghost keeps whole keeps its metadata: all six JPEGs came back with the GPS, and so did a PNG with GPS coordinates and a text comment. In the sample, 24 of 74 PNG pairs had a stored copy the same size as the original, 4 of them wider than 2,000 pixels.

Both GIFs were re-saved, one by 3 bytes and the other by 1.3 percent, and both kept their three frames. Send photographs as JPEG: the Fjord photo stayed 3.5 MB as a PNG and came back at 370 KB as a JPEG.

Ghost's image upload refuses HEIC and AVIF

The image upload takes .jpg, .jpeg, .gif, .png, .svg, .svgz, .ico and .webp (Ghost's config at v6.77.0). Two AVIF files and a real HEIC file all answered HTTP 415 with "Please select a valid image." Convert an iPhone photo with HEIC to JPG first, and send an AVIF as a JPEG.

Visitors get copies made on request, and WebP only on Source's post cards

Ghost cuts the smaller sizes when a browser first asks for them, and cuts them from your original, not from the 2,000-pixel copy: the café photo's 2,400-pixel copy weighed 923,169 bytes against 673,347 for the default image. Ghost's own size list puts copies at 600, 1,000 and 1,600 pixels beside the default image in a post's srcset, and the default theme, Source 1.7.6, adds 320 to 2,000 for the feature image. Our install returned the same bytes for all three Accept headers we tried.

Source's post cards, which sit on its home, tag and author pages and under each post as related posts, ask for WebP. The post's own images don't. In the sample, 22 of 165 image addresses answered WebP and none AVIF.

How we measured

We ran the official ghost:6.77.0 Docker image, the current release when we started, with default configuration, in development mode on SQLite and with no proxy or CDN in front. Ghost 6.78.0 came out the same day, and the image-resizing code and its image settings are unchanged in it. Ghost accepted 70 different files, six of them sent twice, and refused the AVIF and HEIC files. We sent them through the image-upload endpoint of its Admin API and copied the storage folder out afterwards.

Scores are SSIMULACRA 2 from libjxl 0.12.0 of Ghost's copy against a reference: the upload, turned upright where its tag said so and resized to Ghost's size with Pillow 11.1's Lanczos filter, with the two photographs that have a non-sRGB profile converted to sRGB first. The resized arm used ImageMagick 7.1.1-43. Iminify ran through its API on 17 November 2026: Smart with JPG output and a width of 2000, and Lossless with JPG output. MB means a million bytes.

The sample is 60 ghost.io sites from Common Crawl's CC-MAIN-2026-39 index, 165 images, asked for by a script that sent Chrome 157's Accept header and user agent. 59 of the 60 pages reported Ghost 6.78. We didn't upload to Ghost(Pro) and didn't test the editor, a proxy, S3 storage or any other theme.

Questions this post answers

Share this post

Share on X
Share on Facebook
Share on LinkedIn
Share on Reddit
Share on Hacker News
Email
Copy link
Link copied
Bar chart: a copy's bytes as a share of its stored file, median. PNG 14%, JPEG 42%, HEIC 7%, WebP 80%.

Does Substack compress images? Yes: Chrome got lossy WebP copies, a median 22 percent of the file Substack stores

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.

8 min read
Bar chart of bytes against the WebP Framer sends first: the AVIF is 0.695 of the WebP on 111 fresh addresses and 0.709 on 839 replayed ones.

Does Framer compress images? Yes: a WebP first, then an AVIF about 30 percent smaller

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".

7 min read
Bar chart for AVIF files: the median size of Webflow's WebP copy against the AVIF itself, 0.64 at 500 pixels wide, 1.14 at 800, 1.14 at 1080 and 1.45 at 1600.

Does Webflow compress images? Its resized copies are re-encoded, and Webflow says converting to WebP or AVIF is a button

In our sample Webflow didn't turn JPEG and PNG files into WebP or AVIF on its own, but it did re-encode the resized copies it makes. On 15 November 2026 we drew 379 sites at random from Webflow's Made in Webflow showcase, loaded their home pages in Chrome 157 and fetched the Webflow-hosted images that 306 of them loaded. For 300 of 300 image addresses the server sent byte-identical files whether Chrome said it takes AVIF, only WebP, or neither. Nearly all the JPEG copies were libjpeg quality 80, and nearly all the PNG copies were palettes of at most 256 colours. For AVIF files, the median WebP copy at 800 pixels and wider weighed more than the AVIF it came from. One guide to Webflow image optimisation, dated 16 April 2026 on its page, says "Webflow automatically serves WebP versions when browsers support it, but the source images need to be optimized before upload". Webflow's own help says converting to WebP or AVIF is a Compress button in the Assets panel, and that responsive copies are made on upload. We checked the guide's sentence and the help's account of copies against the files.

7 min read

Try it on your own images

Drop a photograph in, pick Smart, and compare the result with the original before you download. Free, no account needed.