Compress Image to 100 KB

Compressing to 100 KB: what the number actually means

A hundred kilobytes is the round number the web has quietly settled on for a good content image. It is large enough for a photograph at full single-column width — around 1000 to 1100 pixels — to look genuinely good, and small enough that it downloads in a fraction of a second on almost any connection. Most performance guidance that specifies a per-image budget lands on or near this figure, and most content management systems that impose a default limit choose something in the same neighbourhood.

It is also the point where compression stops being a compromise and becomes simply good practice. Below 100 KB you are trading quality for weight; at 100 KB, for the majority of ordinary photographs, you are removing bytes nobody would have noticed anyway. The original file straight from a camera carries enormous redundancy: metadata, an oversized colour profile, an embedded preview, and pixel data encoded at a quality far beyond what a screen can show.

The tool searches for the highest quality that fits under 100 KB and only reduces dimensions when it has to, which on most inputs it does not. Expect your original resolution to survive intact on anything up to about two megapixels.

Why people choose the 100 KB target

A hundred kilobytes is chosen more often than any other number in this range, and usually for a good reason: it is roughly where per-image performance guidance lands, roughly where content management systems set their defaults, and roughly where an ordinary photograph at full column width stops being distinguishable from its original. Three independent lines of reasoning converge on the same figure, which is a decent sign that it is the right one.

It is also the number to reach for when someone hands you a slow page and asks you to fix it. Images are almost always the largest share of page weight, and standardising content images at 100 KB typically removes eighty to ninety percent of that weight without a single visible complaint. It is the highest-leverage change available on most sites.

Why a target size beats a quality slider

Every image editor asks you for a quality number and then tells you the file size afterwards. That is backwards for the problem most people actually have, which is a fixed ceiling — an upload form, a page budget, a mailbox limit — and no idea which quality setting will land underneath it. The usual result is a loop: export at 70, check, export at 50, check, export at 60, give up somewhere in the middle.

This tool inverts the relationship. You give it the ceiling and it searches for the quality, using a bisection over the encoder settings that converges in a handful of attempts. That is why it can hit 100 KB reliably on the first pass, and why it can tell you afterwards exactly which quality it used — which is useful information, because it tells you how much headroom you had.

What the tool does not do

It does not upload anything, so it cannot use a server-side encoder like MozJPEG or a full WebP encoder with custom tuning. It uses the encoders built into your browser, which are good but not state of the art. It also does not perform content-aware operations — no smart cropping, no subject detection, no upscaling. And it will not make a compressed file better than its source: compressing an image that has already been through a lossy pass compounds the damage, so always start from the best original you have.

The pipeline, step by step

1. Decode. The file is decoded into a bitmap using the browser image decoder, with EXIF orientation applied so photographs taken sideways come out the right way up. This step also discards every metadata block — EXIF, GPS, camera make and model, colour profile — because a bitmap is just pixels.
2. Resize. If you have set a maximum width or height, the bitmap is drawn into a canvas of the fitted size with the browser high-quality smoothing enabled. Aspect ratio is always preserved; the tool never stretches or crops.
3. Search quality. The canvas is encoded repeatedly at different quality settings, bisecting between the floor you set and the practical ceiling. Each attempt produces a real encoded blob, so the size it reports is the actual file size rather than an estimate. Around eight to ten iterations are enough to converge.
4. Downscale and retry. If even the quality floor exceeds 100 KB, the tool computes a scale factor from the ratio between the achieved size and the target, applies it, and runs the quality search again. It repeats this until the file fits, which for a photograph against a small target may take several rounds.
5. Score fidelity. The original and the result are both drawn to a small common canvas and compared pixel by pixel to produce a peak signal-to-noise ratio. This gives you an objective number rather than a promise.
6. Hand back the blob. The result is offered as a download, individually or bundled into a ZIP that is assembled in the browser.

Why bisection rather than a fixed quality

The relationship between quality setting and file size is not linear and it is not the same for two different images. A photograph of a plain wall at quality 80 might be a tenth the size of a photograph of a forest at the same setting. Any fixed quality number therefore lands wildly differently depending on the picture, which is why exporting at "quality 60" is such an unreliable way to hit a file-size target.

Bisection sidesteps that by measuring rather than predicting: encode at the midpoint, read the real size, move the bracket, repeat until the remaining error is smaller than one quality step.

Why dimensions come after quality

Removing pixels is destructive in a way that lowering quality is not — pixels cannot come back. So the tool exhausts the quality range first and reaches for the scissors only when it must, which is what makes it produce the largest image that fits 100 KB rather than an arbitrarily shrunken one.

What actually fits in 100 KB

The honest answer to "how big an image fits" is that it depends entirely on the picture. Compression works by removing redundancy, and a photograph of a foggy lake has far more redundancy than a photograph of a market crowd. The table below works from a bytes-per-pixel model: for each kind of content, roughly how many bytes a pixel costs at the quality settings this tool typically lands on, and therefore how many pixels a 100 KB budget buys.

Derived from typical encoder output at the quality range this tool settles on: JPEG at roughly 0.34 bytes per pixel for detailed photographs down to 0.09 for flat graphics, with WebP at about 0.65 of the JPEG figure. Dimensions are rounded to the nearest ten pixels. These are planning figures, not guarantees — your own images will land somewhere either side depending on their content.

How to read the table

Find the row that best describes your image and read across. If your photograph is a portrait against a plain studio background, you are in the "ordinary photograph" row and can expect the dimensions shown. If it is a landscape full of trees, you are in the "detailed photograph" row and should expect roughly half the pixel area for the same bytes.

The WebP column is the one to look at if you control where the image is going, because it is the format that will actually be used. The JPEG columns matter when the destination is an upload form, an email or any system whose format handling you cannot predict.

How 100 KB compares with the targets either side

The same photograph at three neighbouring budgets, using the WebP figure for an ordinary photograph. This is the fastest way to tell whether you are on the right tier or one step away from it.

Below this, at 50 KB, you halve the weight and lose about 30% of the linear resolution — often fine for images inside an article. Above, at 150 to 200 KB, you buy the resolution a hero image needs on a retina screen. Very few sites need one number for all three jobs.

Every target in the set: 10 KB  ·  20 KB  ·  30 KB  ·  50 KB  ·  150 KB  ·  200 KB  ·  300 KB  ·  500 KB  ·  1 MB

The single most useful rule

Halving the width and height quarters the pixel count, which roughly quarters the file size at the same quality. That relationship is the lever you have. If a 100 KB target is producing a soft image, the fix is almost always fewer pixels rather than a different encoder or a cleverer setting — and because screens are smaller than camera sensors, those pixels are usually ones nobody would have seen anyway.

What makes an image expensive

• Fine random texture. Foliage, gravel, fabric weave, hair, sand and film grain are incompressible by nature — there is no pattern to exploit.
• High contrast edges. Rendered text, line art and hard shadows force the encoder to spend heavily to avoid visible ringing.
• Noise and heavy sharpening. Both are high-frequency detail, and high-frequency detail is what costs bytes. Denoise before compressing and sharpen lightly, after resizing.

Choosing a format at 100 KB

WebP at 100 KB gives you roughly 1.5 times the pixel area of JPEG for the same bytes, which in practice means either a sharper image at the same dimensions or the same sharpness at around 1300 pixels instead of 1050. Every current browser supports it, and for images served from your own site there is no longer a good reason not to use it.

JPEG remains the safe choice for files that leave your control — attachments, uploads to third-party systems, anything that might be opened by unpredictable software. At 100 KB the compatibility choice costs you very little, because a 100 KB JPEG at 1000 pixels is already a good image.

PNG at 100 KB is genuinely spacious for interface content: a full-width documentation screenshot at 1200 pixels usually fits, and every letter stays sharp. Photographic PNG remains a bad idea at any size.

What "auto" does

With the format set to auto, the tool inspects each image and decides per file. Images with a meaningful alpha channel go to WebP so transparency survives. Images that look photographic go to WebP if your browser can encode it and JPEG otherwise. Images that appear to be flat graphics or screenshots — few distinct colours, large uniform regions — stay as PNG, because forcing them through a lossy encoder would produce a larger and uglier file. If you know what your images are, setting the format explicitly is always more predictable than auto.

The re-encoding trap

Every lossy encode discards information permanently, and the losses compound. A JPEG that has been through three save cycles at quality 80 looks noticeably worse than one saved once at quality 60, even though the last file is larger. This matters here because it is tempting to compress a file, decide it is not small enough, and compress the result again. Do not: change the target and re-run from the original instead. The tool keeps your originals in the list precisely so you can do that.

Transparency and background flattening

JPEG has no alpha channel, so converting a transparent PNG to JPEG has to put something behind it. The tool fills transparent areas with the colour you pick — white by default, which suits most pages but is exactly wrong on a dark background, where it produces a bright rectangle around your logo. If transparency matters, choose WebP or PNG and it survives untouched.

What 100 KB looks like

At 100 KB, ordinary photographs come out essentially indistinguishable from the original at normal viewing distances. The fidelity score typically lands in the high thirties or low forties, a range where side-by-side comparison at 100% zoom is needed to see any difference at all. This is the budget at which you can stop worrying and start standardising.

The exceptions are predictable. Very large images — 3000 pixels wide and up — spread 100 KB thinly and will show softening. Highly detailed subjects consume the budget faster than smooth ones. And anything containing rendered text still belongs in PNG, no matter how generous the byte budget looks, because the problem there is the nature of lossy encoding rather than the amount of data available to it.

If you are setting a house standard, 100 KB for content images and a larger allowance for hero images is a sound and defensible policy. It produces pages that feel fast without anyone in marketing complaining that the photographs look cheap.

A measured result at 100 KB

Vague promises about quality are easy to make, so here is an actual run. A deliberately difficult synthetic photograph — 2000 × 1500 pixels, saturated, and loaded with random noise that no encoder can compress away — was put through this exact tool at the 100 KB target in Chromium. Real photographs are easier than this, so treat these figures as close to a worst case rather than a typical one.

Mistakes people make at 100 KB

• Applying 100 KB to every image on the page. Thumbnails should be far smaller and the hero can afford more. One number for every role is how a page ends up at two megabytes.
• Compressing instead of serving responsive sizes. A phone downloading a desktop-width image wastes most of it. Compression helps; srcset helps more.
• Forgetting that a 3000 px image spreads 100 KB thinly. Cap the width around 1200 to 1600 and the same budget produces a much sharper picture.
• Re-compressing a file that has already been through the process. Losses compound. Change the target and re-run from the original instead.
• Assuming compression fixes Core Web Vitals on its own. It fixes largest contentful paint; layout shift needs explicit width and height in your markup.

Reading the fidelity score

Every result carries a peak signal-to-noise ratio in decibels, computed by comparing the compressed output against the original pixel by pixel. It is a genuine measurement rather than a marketing claim, and it is a good deal more useful than a vague assurance that quality has been preserved.

One caveat worth knowing: peak signal-to-noise ratio measures mathematical difference, not perceived quality. A slight overall brightness shift scores badly while being invisible; blocking concentrated in one corner scores well while being obvious. Use the number as a guide and the comparison slider as the verdict.

What compression artifacts look like

• Blocking. Visible 8 × 8 squares in flat areas, the classic JPEG failure. Appears first in skies and shadows.
• Banding. Smooth gradients breaking into visible steps. Worst in skies, studio backdrops and soft shadow.
• Ringing. Faint halos alongside hard edges, most obvious around text and line art. This is why screenshots belong in PNG.
• Colour bleed. Saturated colour smearing beyond its boundary, because colour is stored at lower resolution than brightness.
• Texture loss. Fine detail flattening into smooth patches — the "waxy" look on skin and fabric.

Getting more quality out of the same budget

1. Crop first. Removing pixels you do not need is free compression, and it concentrates the budget on the part of the image that matters.
2. Cap the dimensions. The most effective single lever. A smaller sharp image beats a larger broken one in almost every context.
3. Use WebP. Roughly a third fewer bytes for the same visual result, and the advantage is widest exactly where budgets are tight.
4. Denoise a noisy original. Sensor noise is random data and random data does not compress. Removing it can cut file size dramatically.
5. Sharpen less. Aggressive sharpening manufactures the high-frequency detail that costs the most bytes.
6. Start from the best original. Compressing an already-compressed file compounds the damage from both passes.

Where 100 KB is the right budget

Choosing a budget by role, not by habit

The most common mistake in image optimisation is applying one number to everything. A page has different kinds of image on it and they deserve different budgets: the hero that determines the largest contentful paint, the content images inside the article, and the thumbnails in a related-posts strip are three different jobs. Give the hero a generous allowance, keep content images moderate, and squeeze thumbnails hard — the thumbnails are where volume multiplies small savings into large ones.

When compression is the wrong answer

• When you should be serving a different size. A phone downloading a desktop-width image wastes most of it. Responsive images through srcset beat compression alone.
• When the image should not be an image. Text rendered into a picture is heavier, unsearchable and inaccessible. Use real text. Charts are often better as SVG.
• When the original is the deliverable. Print files, archival masters and anything a client is paying for should never be replaced by a compressed copy.

Working efficiently

If you are optimising an existing site, start by finding the images that are furthest over budget rather than compressing everything. A handful of five-megabyte hero images usually account for most of a page weight problem, and fixing those five files does more than carefully optimising two hundred thumbnails that were already fine.

Recommended settings for common 100 KB jobs

Starting points rather than rules. The quality floor only sets how far the search may go; the tool still uses the highest setting that fits, so raising the floor makes it give up and downscale sooner rather than producing a bigger file.

1. Test one before running fifty. Drop in a single representative image, look at the result on the comparison slider, and only then commit the batch. Two minutes here saves re-running everything.
2. Decide dimensions before size. Setting a maximum width is what turns a byte budget into a sharp picture. Left unconstrained, a 6000-pixel camera file spends the whole budget on pixels no screen will show.
3. Keep your originals. Compression is one-way. Compress into a new folder and never overwrite the source, because the day you need a print or a different crop will arrive.
4. Group by content type. Photographs, screenshots and logos want different formats; run them as separate batches rather than trusting auto to sort a mixed pile.

Batch processing notes

Every file in a batch is processed independently against the same target, so a mixed set produces mixed results — a simple graphic may land far under the ceiling while a detailed photograph needs downscaling to get there. The results table shows what each file required, which makes the outliers easy to spot and re-run with different settings.

The ZIP is assembled in your browser using stored (uncompressed) entries. That is deliberate: the files inside are already compressed images, so deflating them again would cost time and save almost nothing. The archive opens in every standard tool.

Filenames and organisation

Output files keep the original name with the new extension. If you compress a JPEG to WebP you will get the same base name with a .webp extension, which means a compressed set can sit alongside the originals without collisions. When you are optimising a website this matters: update the references in your markup at the same time, or serve both through a picture element with the WebP first and the JPEG as a fallback.

Accessibility and SEO are separate problems

Compression makes pages faster, and speed helps search ranking indirectly. It does nothing for the things that actually describe your image to a search engine or a screen reader: the alt text, the filename, the caption and the surrounding content. Compress for speed, write proper alt text for everything else, and do not confuse the two.

Image compression terms

Artifact

Visible damage introduced by lossy compression — blocking, banding, ringing or colour bleed.

Banding

Smooth gradients breaking into visible steps. The first failure you will see in skies and studio backdrops.

Blocking

Visible square tiles in flat areas, produced by the block-based transform at the heart of JPEG.

Bytes per pixel

File size divided by pixel count. The single most useful way to reason about whether an image will fit a byte budget.

EXIF

Metadata a camera embeds — exposure, lens, timestamp and often GPS coordinates. Removed entirely by re-encoding.

Lossless

Compression that reconstructs the original exactly. PNG and lossless WebP; larger files, no artifacts.

Lossy

Compression that discards information permanently in exchange for much smaller files. JPEG, lossy WebP, AVIF.

LCP

Largest Contentful Paint, the Core Web Vitals metric measuring when the biggest element becomes visible. Usually an image.

PSNR

Peak signal-to-noise ratio, the fidelity score shown for each result. Higher is closer to the original.

Quality setting

The 0 to 100 parameter given to a lossy encoder. It is not a percentage of anything and the same number means different things in different encoders.

Re-encoding

Decoding an image and encoding it again. Each lossy round trip compounds the damage of the last.

Ringing

Faint halos beside hard edges, most visible around text. The reason screenshots belong in PNG.

WebP

An image format supporting both lossy and lossless compression with transparency, typically 25 to 35% smaller than JPEG.

Frequently asked questions