Skip to content
TeardownBy 8 min read

Teardown: a Wikipedia article, full page

One public page, 27,404 px tall, captured with plain Chromium. Five things broke. Each has a crop of the evidence and a fix.

TeardownCaptured 12 November 2026Chromium 1531280 x 800

The page is the English Wikipedia article “Octopus”[1], last edited 26 September 2026. It was picked because anyone can open it and because it is long: a long, plain-looking page is what most captures are. The crops below are cut from the capture to show the layout; the credits are at the end.

Breaksaround 3,914 px from the top

Pictures are empty boxes

A stretch of the article in a plain capture: a framed empty box with the caption 'Eye of common octopus' and no picture in it.
Plain capture. The box has its frame and its caption, and no picture. About 4,950 px down.
The same stretch of the article after the pictures were loaded: the box now holds a photograph.
The same place after the pictures were loaded. Same text, same height.

The page has 55 img elements. 30 of them are article pictures that show somewhere in the page. 52 of the 55 carry loading="lazy", so the browser fetches one only when it comes near the window. The load event does not wait for them: the page counts as loaded while pictures within sight can still be missing.[2] A plain full-page capture then draws what is there, and an empty box where a picture will be.

CapturePictures drawn
Plain page.screenshot, nothing else4 of 30
Steps of 720 px, 90 ms apart, back to the top, then wait up to 4 s30 of 30
Every loading="lazy" switched to eager, then wait for the pictures30 of 30
The CurlShot renderer scroll, run in plain Playwright30 of 30

In the plain capture the empty boxes lay between 3,914 px and 16,945 px. After the stepped scroll the lab still counted 4 unloaded thumbnails, but they sit inside the folded boxes at the foot of the page. The stepped, the eager and a dark-scheme capture came out as the same file (sha-256 starts 1a6183b59d91), so none of the 4 shows in a picture. The renderer scroll took 2.8 s with a limit of 20,000 px and 4.1 s with 30,000 px.

The fix

Load the pictures before the picture is taken: switch loading="lazy" to eager and wait until every picture is complete, or scroll through the page in steps smaller than the window. CurlShot does both unless full_page_scroll is off. How the scroll copes with pages that load pictures in other ways is in lazy images and full-page captures, measured.
Shiftsaround 26,740 px from the top

The page changes height after it loads

Read at the load event, the page was 29,304 px tall in 5 of 6 loads. 500 ms later it was 27,404 px in all 6. The 1,900 px are at the foot: five navigation boxes are open when the page arrives and fold shut a moment later.

BoxAt loadFolded
Navigation box 1561 px36 px
Navigation box 2522 px36 px
Navigation box 3834 px36 px
Navigation box 4168 px168 px
Navigation box 5127 px36 px
Together2,212 px312 px

A script that reads the height at load and then sets a window or a clip of that size draws 1,900 px of empty page. Plain Chromium measured when it took the picture, so the captures here are 27,404 px.

The fix

Wait until the height reads the same twice, half a second apart. The delay option of CurlShot adds a fixed wait after the load, up to 30 seconds.
The foot of the settled page: the External links list, then five navigation boxes each folded to a single title row, then the page footer.
The foot of the settled page, from 26,500 px down: five navigation boxes, each folded to a title bar.
Shiftsaround 250 px from the top

The contents list moves into the article

The top of a capture taken without a scroll: the contents list stands in the left margin and the Appearance panel in the right margin, next to the first paragraphs.
Capture with no scroll. The contents list and the Appearance panel stand at the top, where a reader finds them.
Part of the article about 11,000 px down, with the contents list in the left margin and the Appearance panel in the right margin beside a paragraph on defence.
Capture after a scroll to 9,000 px. The same two panels, now 11,117 px down in the middle of the article.

The contents list and the Appearance panel use position: sticky: they stay at a set distance from the top of the window while the page moves under them.[3] A full-page capture is one tall picture of the page at one moment. A sticky element is drawn where the window was at that moment, once.

Scroll before the captureScroll after the captureContents list found at
0 px0 px143 px
4,000 px4,884 px5,025 px
9,000 px10,976 px11,117 px
20,000 px20,985 px21,126 px

Each picture held one contents list, at the place the window had reached, and none at the top. The scroll position also moved during the capture (from 4,000 to 4,884 px in the second row), so the list does not land on the exact scroll value.

The fix

Scroll back to the top, wait a moment, then take the picture. The CurlShot renderer ends its scroll that way.
The bottom edge of a capture cut at 20,000 px: a numbered reference list that stops in the middle of an entry.
The last 700 px of a capture cut at 20,000 px. It stops inside an entry of the reference list.
Cutsaround 20,000 px from the top

The default height cap cuts the page

The page is 27,404 px tall. The default full_page_max_height of CurlShot is 20,000 px, and the largest value is 30,000 px. At the default, the capture keeps 73% of the page and ends inside the references, whose heading is at 17,763 px.

20,000 px kept, 7,404 px lost: the end of the references, the External links section (at 26,541 px) and the foot.

Size grows too. The whole page as a PNG is 9.2 MB and took 7.8 s to make; as a JPEG at quality 80 it is 5.5 MB. A longer article, “Moon”, was 63,945 and 67,337 px tall in two loads. Plain Chromium returned no full-page picture of it within 30 s on either load, and none within 150 s on a retry of the second one.

The fix

Set full_page_max_height to 30,000 for a page like this. Past 30,000 px take slices with clip_y and clip_height, or capture the one section that matters with a selector.
Dependsaround 90 px from the top

The fonts are the host’s fonts

The top of the article in the fonts of the capture host: a serif heading and a sans-serif body.
The fonts the host chose: Liberation Serif for headings, Liberation Sans for text.
The same top of the article with every font swapped for DejaVu Sans: wider letters and an earlier line break.
The same page with every font swapped for DejaVu Sans, also installed on the host.

The page names no web font. The headings ask for "Linux Libertine", Georgia, Times, "Source Serif 4", serif and the text for sans-serif. The host had none of the named fonts, so Chromium drew Liberation Serif and Liberation Sans. With every font on the page swapped for DejaVu Sans, the height went from 27,404 px to 30,155 px: 2,751 px more, because wider letters wrap into more lines.

The swap is a stand-in for another machine, not a test of one. What it shows is that this page's line breaks, and so its height, belong to the machine that drew it.

The fix

Make the captures you compare on one machine, or one service. Captures of the same address from a laptop and from a server will differ in height and line breaks.

What was looked for and not found

A teardown also says what held. These checks came back clean on this page, for this visit.

  • CleanConsent banner

    No dialog or modal on the page. The one element with "cookie" in its name is the footer link "Cookie statement". No site notice.

  • CleanCarousel

    No element with a carousel, slider or swiper class. The two image strips are plain pictures side by side.

  • CleanAutoplay

    Two video elements, both with preload none and no autoplay. Their poster pictures were drawn in the plain capture.

  • CleanFrames and galleries

    No iframe and no gallery element.

  • CleanDark mode

    A capture with the colour scheme set to dark came out byte for byte the same as the light one. Readers who are not signed in get the light theme.

The consent check is for this server's address on 12 November 2026; a visit from another country may see something else. A page that does show a banner is the job of the screenshot without cookie banners tool, and of the block_cookie_banners option.

The three fixes in one script

Plain Playwright. It loads the pictures, waits for a steady height and scrolls back to the top. Run on this page it gave a 27,404 px picture in 9.1 s from the first request to the file, and the same file as the eager capture above (sha-256 starts 1a6183b59d91).

capture.mjs
const url = 'https://en.wikipedia.org/wiki/Octopus'
const browser = await chromium.launch()
const page = await browser.newPage({ viewport: { width: 1280, height: 800 } })
await page.goto(url, { waitUntil: 'load' })

// 1. Lazy pictures: load them all now.
await page.evaluate(() =>
  document.querySelectorAll('img[loading="lazy"]').forEach((i) => { i.loading = 'eager' }))
await page.waitForFunction(() => [...document.images].every((i) => i.complete), null, { timeout: 10000 })

// 2. Height: wait until it stops moving.
let last = 0
for (;;) {
  const h = await page.evaluate(() => document.documentElement.scrollHeight)
  if (h === last) break
  last = h
  await page.waitForTimeout(500)
}

// 3. Sticky panels: back to the top before the picture.
await page.evaluate(() => scrollTo(0, 0))
await page.screenshot({ path: 'page.png', fullPage: true })
await browser.close()

The same capture as one request. It was not sent in the lab; its options are read from the option reference.

cURL
curl -G "https://curlshot.com/api/v1/screenshot" \
  --data-urlencode "access_key=YOUR_ACCESS_KEY" \
  --data-urlencode "url=https://en.wikipedia.org/wiki/Octopus" \
  --data-urlencode "full_page=true" \
  --data-urlencode "full_page_max_height=30000" \
  --data-urlencode "delay=2" \
  --data-urlencode "format=jpeg" \
  --output octopus.jpg

full_page_scroll is on by default, so the pictures load as in the first section. One successful render is one screenshot whatever its length; see pricing for the free plan.

How it was run

On 12 November 2026. The English Wikipedia article "Octopus" was opened in a new Chromium 153.0.8010.12 context, 1280 x 800, through Playwright 1.63.0 on a Linux server, and captured with page.screenshot({ fullPage: true }). Each variant used its own new context. Pictures were counted from the page: an article picture is an img in the article body that is at least 60 px wide, and it counts as drawn when it has finished loading.

The plain, stepped, eager, dark-scheme and sticky captures are single captures; the height at load comes from 6 loads. The renderer scroll is the function of the CurlShot renderer, read from its source file and run unchanged inside plain Playwright. No call to the API was sent. One page on one day shows what Chromium did with that page, not with yours.

Credits: the text of Wikipedia is available under CC BY-SA 4.0 and its pictures carry their own licences.[4] The crops are small cuts of the capture, shown to explain the layout. The full article, its authors and the licence of each picture are on Wikipedia.

Related reading

For the many other reasons a capture differs from your own browser, read why website screenshots look wrong. For the scroll step by step, read lazy images and full-page captures, measured.

Questions people ask

Why are the pictures empty boxes in my full-page screenshot?

The page marks them loading="lazy", and the load event fires before lazy pictures arrive. In the test, a plain capture drew 4 of 30. Scrolling in steps of 90% of the window, or switching the pictures to eager, drew all 30. CurlShot scrolls through the page first unless full_page_scroll is off.

Why is my full-page screenshot taller or shorter than the page I see?

The page may still be changing. Here the height at the load event was 29,304 px and 27,404 px after 500 ms, because open navigation boxes folded shut. Wait until the height reads the same twice. A different font on the capture host can also move it: 27,404 px became 30,155 px with another font.

Why does a menu or contents list appear in the middle of the capture?

It is position: sticky or fixed and the page was scrolled when the picture was taken. In the test the contents list was drawn at 11,117 px after a scroll to 9,000 px. Scroll back to the top first; the CurlShot renderer does that at the end of its scroll.

How tall can a full-page capture be?

On CurlShot the default cap is 20,000 px and the largest full_page_max_height is 30,000 px. A page taller than that needs slices with clip_y and clip_height. In plain Chromium a 63,945 px article returned no picture within 30 s.

May I capture and show a Wikipedia page?

Wikipedia text is available under CC BY-SA 4.0, and its pictures carry their own licences, so a reuse needs credit and a look at each picture's file page. This is a pointer to the reuse guide, not legal advice.

Sources

  1. [1]Wikipedia, the free encyclopedia, Octopus. Checked .
  2. [2]MDN Web Docs, Lazy loading. Checked .
  3. [3]MDN Web Docs, position: sticky. Checked .
  4. [4]Wikipedia, the free encyclopedia, Reusing Wikipedia content. Checked .

Screenshot any page in one step

Paste an address into CurlShot and download the screenshot. A free account gives you 100 screenshots every month. No card needed.