Pick the smallest label a reader needs to see. Export AVIF, WebP and JPEG under the same byte cap and display each one at the card width. Look at the label before comparing larger areas such as photographs.
Start with one uncompressed or losslessly saved screenshot. Make a working copy at the target pixel dimensions and use it as the source for every encoder. If you test versions made from earlier compressed files, they start with different amounts of detail.
Set a budget in bytes
For an illustrative 1,200 by 750 pixel screenshot, choose a cap of 75 KiB, or 76,800 bytes. Export one AVIF, one WebP and one JPEG. Change each encoder's quality setting until the candidate comes near the budget without going over. The encoders use their quality scales differently. Treat 65 in one format as its own setting. Compare the actual byte counts from your file manager or use wc -c on the three files.
Keep the source crop, pixel dimensions and color conversion fixed. If the screenshot contains transparency, flatten the master to the intended background before this three-way test. JPEG cannot preserve the screenshot's alpha channel. If preserving transparency is a product requirement, run a separate comparison among formats that support it. MDN's image format guide summarizes these constraints, and Google's WebP documentation describes its lossy and lossless modes.
Score the parts people need to read
| Region | Look for |
|---|---|
| Small black menu label | Missing strokes and pale ringing around letters |
| Thin colored line on a gray background | Color bleed beyond the edge |
| Flat panel or long gradient | Blocks, mottling or steps between tones |
| Text in a selected or disabled row | A label whose contrast changed enough to hinder reading |
Compare each candidate at the card's real width on a phone and a desktop. Then inspect the native-size export to see where problems appeared. If a label is already unreadable in the master when displayed at the card width, enlarge the interface or crop to the important panel before retesting. Additional image quality cannot restore a label that the layout renders too small.
Record the format, output bytes and whether each target label is readable. If none of the lossy candidates survives, try lossless WebP or PNG and compare its measured size with your budget. For web delivery, test the actual browser and offer a fallback if you use an encoding absent from a required client. In a test page, temporarily remove the preferred source for the final check and confirm the fallback decodes in a client you need to support.
Recreate the eventual viewing conditions
Pick the screenshot section a reader needs to understand: the name of a setting, for example, and the state indicated by a small icon beside it. Save a lossless copy of the screenshot before cropping. Make the crop so both elements remain visible at the target display size. Then use that same raster as the source for each encoding. Keep output pixel dimensions independent of the CSS size in the page. A candidate can be 1,200 pixels wide while the browser fits it into a much smaller card.
Try the screenshot in a card that matches the actual layout. If another title or button is next to the screenshot, include it in the proof. A small screenshot that makes sense on a blank page may be hard to read beside stronger text and higher contrast elements. If the reader needs to see only one panel, test a crop or a capture with a larger interface scale before making a bigger file.
Take one image with mostly solid panels and another with a photo or gradient if your actual pages use both. Efficiency at the same display size can depend on content. A WebP setting that works for a photograph inside an app may soften tiny labels elsewhere. Reuse the same test caps across the candidates for each screenshot. Record which format and encoding mode you used, since a lossless result and a low-quality lossy result answer different questions.
Compare options near the cap
Start with a few settings high enough to look plausible and adjust when a file is too large. Report the byte count rather than assuming an equal quality slider value means an equal image. For a worked test at 76,800 bytes, imagine one candidate is only 35,000 bytes and already passes. You can keep it that small or try a sharper setting while staying under the limit if that extra detail matters. Another candidate may remain larger than the limit even when it loses the lines readers need. Record that result as a limitation for this screenshot.
Inspect the same crop at both actual size and high zoom. At actual size, have a second reader identify the label without the original beside it. At high zoom, check whether a pale line around the letters came from the encoder or was part of the source. Compare against the uncompressed master at the same scale. Keep sample coordinates and the setting names so a later run uses the same test region.
Score each visible requirement independently. If a file keeps the text readable and has some banding in a decorative panel, it might be acceptable. If the screenshot illustrates a gradient tool, that same banding may change the point of the image. In a performance test, verify the chosen file actually travels through the intended CDN or image pipeline. The server can resize or re-encode it after you upload it.
Deliver a reliable fallback when one is needed
Open the served page and check which image file the browser requests at the intended device width. Compare its reported size and crop with the file you approved. If the platform serves different image sources per browser, inspect those variants too. Give the fallback the same subject and use it in the documented client that needs it. Test the real URL path so a fallback cannot silently fail to load in production.
Keep the source image, the acceptable format settings and the shortest readable text you used for the check. On a different card shape, the informative label can be cut before compression matters. Use the three-shape crop exercise when the same screenshot must remain recognizable in several layouts. Then repeat the byte-budget test on the crop that actually ships.


