CSS Image Values and Replaced Content Module Level 4

W3C Working Draft 11 September 2012

This Version:
Latest Version:
Editor's Draft:
Issue Tracking:
www-style@w3.org with subject line “[css4-images] … message topic …
Elika J. Etemad (Mozilla)
Tab Atkins Jr. (Google)


CSS is a language for describing the rendering of structured documents (such as HTML and XML) on screen, on paper, in speech, etc. This module contains the features of CSS level 4 relating to the <image> type and replaced elements. It includes and extends the functionality of CSS Image Values and Replaced Content Module Level 3. The main extensions compared to level 3 are several additions to the ‘<image>’ type, additions to the ‘<gradient>’ type, extensions to the ‘image()’ function, definitions for interpolating several ‘<image>’ types, and several properties controlling the interaction of replaced elements and CSS's layout models.

Status of this document

This section describes the status of this document at the time of its publication. Other documents may supersede this document. A list of current W3C publications and the latest revision of this technical report can be found in the W3C technical reports index at http://www.w3.org/TR/.

Publication as a Working Draft does not imply endorsement by the W3C Membership. This is a draft document and may be updated, replaced or obsoleted by other documents at any time. It is inappropriate to cite this document as other than work in progress.

The (archived) public mailing list www-style@w3.org (see instructions) is preferred for discussion of this specification. When sending e-mail, please put the text “css4-images” in the subject, preferably like this: “[css4-images] …summary of comment…

This document was produced by the CSS Working Group (part of the Style Activity).

This document was produced by a group operating under the 5 February 2004 W3C Patent Policy. W3C maintains a public list of any patent disclosures made in connection with the deliverables of the group; that page also includes instructions for disclosing a patent. An individual who has actual knowledge of a patent which the individual believes contains Essential Claim(s) must disclose the information in accordance with section 6 of the W3C Patent Policy.

This is the first public working draft of css4-images.

Table of contents

1. Introduction

This section is not normative.

In CSS Levels 1 and 2, image values, such as those used in the ‘background-image’ property, could only be given by a single URL value. This module introduces additional ways of representing 2D images, for example as a list of URIs denoting fallbacks, as a reference to an element, or as a gradient.

This module also defines several properties for manipulating raster images and for sizing or positioning replaced elements such as images within the box determined by the CSS layout algorithms. It also defines in a generic way CSS's sizing algorithm for images and other replaced elements.

1.1. Module Interactions

This module defines and extends the ‘<image>’ value type defined in [CSS3VAL]. Furthermore it replaces the ‘<url>’ type in the ‘background-image’ and ‘list-style-image’ definitions in CSS1 and CSS2 and adds ‘<image>’ as an alternative to ‘<url>’ in the ‘content’ property's value. It is presumed that CSS specifications beyond CSS2.1 will use the ‘<image>’ notation in place of ‘<url>’ where 2D images are expected. (See e.g. [CSS3BG].)

Of the properties defined in this module, only ‘image-resolution’ applies to ::first-line and ::first-letter.

1.2. Values

This specification follows the CSS property definition conventions from [CSS21]. Value types not defined in this specification are defined in CSS Level 2 Revision 1 [CSS21]. Other CSS modules may expand the definitions of these value types: for example [CSS3COLOR], when combined with this module, expands the definition of the <color> value type as used in this specification.

In addition to the property-specific values listed in their definitions, all properties defined in this specification also accept the global keywords as their property value. For readability they has not been repeated explicitly.

2. Planned Features

  1. Resolution tagging in image()
  2. Logical keywords for specifying linear gradient direction.
  3. Conical gradients
  4. More interpolation modes for gradients: http://lists.w3.org/Archives/Public/www-style/2012Jan/0635.html
  5. How to transition/animate gradients
  6. Ability to reference an SVG paint server via url (maybe as element(<string>)?)

3. Image Values: the <image> type

The ‘<image>’ value type denotes a 2D image. It can be a url reference, image notation, or gradient notation. Its syntax is:

<image> = <url> | <image-list> | <image-set> | <gradient>

An ‘<image>’ can be used in many CSS properties, including the ‘background-image’, ‘list-style-image’, ‘cursor’ properties [CSS21] (where it replaces the ‘<url>’ component in the property's value).

In some cases, an image is invalid, such as a ‘<url>’ pointing to a resource that is not a valid image format. An invalid image is rendered as a solid-color ‘transparent’ image with no intrinsic dimensions. However, invalid images have special behavior in some contexts, such as the ‘image()’ notation.

3.1. Image References and Image Slices: the ‘<url>’ type and ‘url()’ notation

The simplest way to indicate an image is to reference an image file by URL. This can be done with the url()’ notation, defined in [CSS21].

In the example below, a background image is specified with ‘url()’syntax:

background-image: url(wavy.png);

If the UA cannot download, parse, or otherwise successfully display the contents at the URL as an image, it must be treated as an invalid image.

3.2. Image Fallbacks and Annotations: the ‘image()’ notation

The ‘image()’ function allows an author to:

The ‘image()’ notation is defined as:

<image-list> = image( <image-tags>? [ <image-src> , ]* [ <image-src> | <color> ] )
<image-tags> = [ ltr | rtl ] <image-src> = [ <url> | <string> ]

Each ‘<string>’ or ‘<url>’ inside ‘image()’ represents an image, just as if the url()’ notation had been used. As usual for URLs in CSS, relative URLs are resolved to an absolute URL (as described in Values & Units [CSS3VAL]) when a specified ‘image()’ value is computed.

3.2.1. Image Fragments

When a URL specified in ‘image()’ represents a portion of a resource (e.g. by the use of media fragment identifiers) that portion is clipped out of its context and used as a standalone image.

For example, given the following image and CSS: [9 circles, with 0 to 8 eighths filled in]

background-image: image('sprites.svg#xywh=40,0,20,20')

...the background of the element will be the portion of the image that starts at (40px,0px) and is 20px wide and tall, which is just the circle with a quarter filled in.

So that authors can take advantage of CSS's forwards-compatible parsing rules to provide a fallback for image slices, implementations that support the ‘image()’ notation must support the xywh=#,#,#,# form of media fragment identifiers for images specified via ‘image()’. [MEDIA-FRAGS]

Note that image fragments can also be used with the ‘url()’ notation. However, a legacy UA that doesn't understand the media fragments notation will ignore the fragment and simply display the entirety of the image.

Since the ‘image()’ notation requires UAs to support media fragments, authors can take advantage of CSS's forward-compatible parsing rules to provide a fallback when using an image fragment URL:

background-image: url('swirl.png'); /* old UAs */
background-image: image('sprites.png#xywh=10,30,60,20'); /* new UAs */

3.2.2. Image Fallbacks

Multiple ‘<image-src>s’ can be given separated by commas, in which case the function represents the first image that's not an invalid image. The final argument can specify a ‘<color>’ to serve as an ultimate fallback; this can be used, e.g. for ‘background-image’, to ensure adequate contrast if none of the preceding ‘<image-src>s’ can be used. If the final argument is a ‘<color>’, it represents a solid-color image of the given color with no intrinsic dimensions. If all of the provided ‘<image-src>s’ are invalid images and a fallback color was not provided as the last argument, the entire ‘image()’ function must be treated as an invalid image.

The rule below would tell the UA to load ‘wavy.svg’ if it can; failing that to load ‘wavy.png’; and failing that to display ‘wavy.gif’. For example, the browser might not understand how to render SVG images, and the PNG may be temporarily 404 (returning an HTML 404 page, which the browser can't decode as an image) due to a server move, so the GIF is used until one of the previous problems corrects itself.

background-image: image("wavy.svg", 'wavy.png' , "wavy.gif");

The fallback color can be used to ensure that text is still readable even when the image fails to load. For example, the following code works fine if the image is rectangular and has no transparency:

body      { color: black; background: white; }
p.special { color: white; background: url("dark.png") black; }

When the image doesn't load, the background color is still there to ensure that the white text is readable. However, if the image has some transparency, the black will be visible behind it, which is probably not desired. The ‘image()’ function addresses this:

body      { color: black; background: white; }
p.special { color: white; background: image("dark.png", black); }

Now, the black won't show at all if the image loads, but if for whatever reason the image fails, it'll pop in and prevent the white text from being set against a white background.

If a URL uses a fragment identifier syntax that the implementation does not understand, or does not consider valid for that type of image, the URL must be treated as representing an invalid image. This error-handling is limited to image(), and not in the definition of URL, for legacy compat reasons.

For example, if a future specification defined a way to refer to a specific frame of an animated GIF with a fragment identifier, an author could write the following to get newer browsers to use the GIF's frame, and older browsers to instead download the fallback image:

background-image: image('cat_meme.gif#frame=5', 'lolcat.png');

3.2.3. Solid-color Images

At times, one may need a solid-color image for a property or function that does not accept the ‘<color>’ type directly. The ‘image()’ function can be used for this: by specifying only a color without any URLs, the function immediately falls back to representing a solid-color image of the chosen color.

background-image: image(rgba(0,0,255,.5)), url("bg-image.png");

In the above, the background is the image "bg-image.png", overlaid with partially-transparent blue.

3.2.4. Bidi-sensitive Images

Before listing any <image-src>s, the author may specify a directionality for the image, similar to adding a dir attribute to an element in HTML. If a directional image is used on or in an element with opposite direction, the image must be flipped in the inline direction (as if it was transformed by, e.g., scaleX(-1), if the inline direction is the X axis).

Note that, absent this declaration, images default to no directionality at all, and thus don't care about the directionality of the surrounding element.

A list may use an arrow for a bullet that points into the content. If the list can contain both LTR and RTL text, though, the bullet may be on the left or the right, and an image designed to point into the text on one side will point out of the text on the other side. This can be fixed with code like:

<ul style="list-style-image: image(ltr 'arrow.png');">
	<li dir='ltr'>My bullet is on the left!</li>
	<li dir='rtl'>MY BULLET IS ON THE RIGHT!</li>

This should render something like:

⇒ My bullet is on the left!

In LTR list items, the image will be used as-is. In the RTL list items, however, it will be flipped in the inline direction, so it still points into the content.

3.3. Resolution Negotiation: the ‘image-set()’ notation

Delivering the most appropriate image resolution for a user's device can be a difficult task. Ideally, images should be in the same resolution as the device they're being viewed in, which can vary between users. However, other factors can factor into the decision of which image to send; for example, if the user is on a slow mobile connection, they may prefer to receive lower-res images rather than waiting for a large proper-res image to load. The ‘image-set()’ function allows an author to ignore most of these issues, simply providing multiple resolutions of an image and letting the UA decide which is most appropriate in a given situation.

This solution assumes that resolution is a proxy for filesize, and therefore doesn't appropriately handle multi-resolution sets of vector images, or mixing vector images with raster ones (e.g. for icons). For example, use a vector for high-res, pixel-optimized bitmap for low-res, and same vector again for low-bandwidth (because it's much smaller, even though it's higher resolution).

The syntax for ‘image-set()’ is:

<image-set> = image-set( [ <image-set-decl>, ]* [ <image-set-decl> | <color>] )
<image-set-decl> = [ <image> | <string> ] <resolution>

The ‘image-set()’ function can not be nested inside of itself, either directly or indirectly (as an argument to another <image> type).

WebKit's experimental implementation has an ‘x’ unit which is identical to ‘dppx’. Should we add it?

Each ‘<string>’ or ‘<url>’ inside ‘image-set()’ represents an image, just like in ‘image()’.

Every <image-set-decl> in a given ‘image-set()’ must have a different <resolution>, or else the function is invalid.

The value that a given ‘image-set()’ represents is determined by the UA first ordering the set of <image-set-decl>s in a UA-determined order (which should be based on the <resolution> of each <image-set-decl>, but may include other factors that the UA deems relevant, such as connection speed). The function then represents the first such image that is not an invalid image.

If a fallback color was provided, the UA may decide at any point that the remaining <image-set-decl>s are unsuitable, and choose to use the fallback color instead, in which case the function represents a solid-color image of the given color with no intrinsic dimensions.

If all of the <image-set-decl>s represent invalid images, then if a fallback color was provided, the function represents a solid-color image of the given color with no intrinsic dimensions; otherwise, the function represents an invalid image.

HTML currently doesnt' do fallback if the first-chosen image doesn't load. We should harmonize with HTML, either by convincing them to do fallback, or by removing our own fallback.

If an <image-set-decl> is chosen and successfully loaded, the associated <resolution> is the image's intrinsic resolution.

It's been suggested that we don't need a full <resolution> here, and should instead just use the ‘x’ unit by itself. This seems silly to me, since the ‘x’ unit is just an easier synonym for ‘dppx’, but thoughts? Beyond theoretical purity considerations, it seems that ‘dpi’ can be useful for use-cases involving printing, where dpi is actually used to talk about images and having an integer correlation between dots and pixels is either unnecessary or strictly non-sensical (as the screening resolution can depend on the colors being used).

This example shows how to use ‘image-set()’ to provide an image in three versions: a "normal" version, a "high-res" version, and an extra-high resolution version for use in high-quality printing (as printers can have extremely high resolution):

background-image: image-set( "foo.png" 1x,
                             "foo-2x.png" 2x,
                             "foo-print.png" 600dpi );

3.4. Using Elements as Images: the ‘element()’ notation

The ‘element()’ function allows an author to use an element in the document as an image. As the referenced element changes appearance, the image changes as well. This can be used, for example, to create live previews of the next/previous slide in a slideshow, or to reference a canvas element for a fancy generated gradient or even an animated background.

Note that the ‘element()’ function only reproduces the appearance of the referenced element, not the actual content and its structure. Authors should only use this for decorative purposes, and must not use ‘element()’ to reproduce an element with significant content across the page. Instead, just insert multiple copies of the element into the document.

The syntax for ‘element()’ is:

<element-image> = element( <id-selector> )

where ‘<id-selector>’ is an ID selector [SELECT].

Need to add some syntax + API for referring to an element that's just in script, not in the document, so you can, say, make an HTML <canvas> and just animate it directly.

Do we need to be able to refer to elements in external documents (such as SVG paint servers)? Or is it enough to just use url() for this?

This name conflicts with a somewhat similar function in GCPM. This needs to be resolved somehow.

Want the ability to do "reflections" of an element, either as a background-image on the element or in a pseudo-element. This needs to be specially-handled to avoid triggering the cycle-detection.

When we have overflow:paged, how can we address a single page in the view?

The ‘element()’ function references the element matched by its argument. If multiple elements are matched, the function references the first such element.

The image represented by the ‘element()’ function can vary based on whether the element is visible in the document:

an element that is rendered and is not a descendant of a replaced element

The function represents an image with its intrinsic size equal to the decorated bounding box of the referenced element:

Note: Because images clip anything outside their bounds by default, this means that decorations that extend outside the decorated bounding box, like box shadows, may be clipped.

The image is constructed by rendering the referenced element and its descendants (at the same size that they would be in the document) over an infinite ‘transparent’ canvas, positioned so that the edges of the decorated bounding box are flush with the edges of the image.

Requiring some degree of stacking context on the element appears to be required for an efficient implementation. Do we need a full stacking context, or just a pseudo-stacking context? Should it need to be a stacking context normally, or can we just render it as a stacking context when rendering it to element()?

If the referenced element has a transform applied to it or an ancestor, the transform must be ignored when rendering the element as an image. [CSS3-TRANSFORMS]

If the referenced element is broken across pages, the element is displayed as if the page content areas were joined flush in the pagination direction, with pages' edges corresponding to the initial containing block's start edge aligned. Elements broken across lines or columns are just rendered with their decorated bounding box.

Implementations may either re-use existing bitmap data generated for the referenced element or regenerate the display of the element to maximize quality at the image's size (for example, if the implementation detects that the referenced element is an SVG fragment); in the latter case, the layout of the referenced element in the image must not be changed by the regeneration process. That is, the image must look identical to the referenced element, modulo rasterization quality.

As a somewhat silly example, a <p> element can be reused as a background elsewhere in the document:

#src { color: white; background: lime; width: 300px; height: 40px; }
#dst { color: black; background: element(#src); padding: 20px; margin: 20px 0; }
<p id='src'>I'm an ordinary element!</p>
<p id='dst'>I'm using the previous element as my background!</p>
an element that is not rendered, but which provides a paint source

The function represents an image with the intrinsic size and appearance of the paint source. The host language defines the size and appearance of paint sources.

For example, the ‘element()’ function can reference an SVG <pattern> element in an HTML document:

<!DOCTYPE html>
		<pattern id='pattern1'>
			<path d='...'>
<p style="background: element(#pattern1)">
	I'm using the pattern as a background!
	If the pattern is changed or animated, 
	my background will be updated too!

HTML also defines that a handful of elements, such as <canvas>, <img>, and <video>, provide a paint source. This means that CSS can, for example, reference a canvas that's being drawn into, but not displayed in the page:

<!DOCTYPE html>
	var canvas = document.querySelector('#animated-bullet');
	canvas.width = 20; canvas.height = 20;
<canvas id='animated-bullet' style='display:none'></canvas>
<ul style="list-style-image: element(#animated-bullet);">
	<li>I'm using the canvas as a bullet!</li>
	<li>So am I!</li>
	<li>As the canvas is changed over time with Javascript,
	    we'll all update our bullet image with it!</li>
anything else

The function represents an invalid image.

For example, all of the following ‘element()’ uses will result in a transparent background:

<!DOCTYPE html>
<p id='one' style="display:none;">one</p>
<iframe src="http://example.com">
	<p id='two'>I'm fallback content!</p>
	<li style="background: element(#one);">
	  A display:none element isn't rendered, and a P element
	  doesn't provide a paint source.
	<li style="background: element(#two);">
	  The descendants of a replaced element like an IFRAME
	  can't be used in element() either.
	<li style="background: element(#three);">
	  There's no element with an id of "three", so this also
	  gets rendered as a transparent image.

An element is not rendered if it does not have an associated box. This can happen, for example, if the element or an ancestor is ‘display:none’. Host languages may define additional ways in which an element can be considered not rendered; for example, in SVG, any descendant of a <defs> element is considered to be not rendered.

The ‘element()’ function can be put to many uses. For example, it can be used to show a preview of the previous or next slide in a slideshow:

<!DOCTYPE html>
function navigateSlides() {
	var currentSlide = ...;
	document.querySelector('#prev-slide').id = '';
	document.querySelector('#next-slide').id = '';
	currentSlide.previousElementSibling.id = 'prev-slide';
	currentSlide.nextElementSibling.id = 'next-slide';
#prev-preview, #next-preview { 
	position: fixed;
#prev-preview { background: element(#prev-slide); }
#next-preview { background: element(#next-slide); }
<a id='prev-preview'>Previous Slide</a>
<a id='next-preview'>Next Slide</a>
<section class='slide'>...</section>
<section class='slide current-slide'>...</section>

In this example, the navigateSlides function updates the ids of the next and previous slides, which are then displayed in small floating boxes alongside the slides. Since you can't interact with the slides through the ‘element()’ function (it's just an image), you could even use click handlers on the preview boxes to help navigate through the page.

3.4.1. Paint Sources

Host languages may define that some elements provide a paint source. Paint sources have an intrinsic width, height, and appearance, separate from the process of rendering, and so may be used as images even when they're not rendered.

In HTML, the <img>, <video>, and <canvas> elements provide paint sources (defined in each element's section in HTML5).

In SVG, any element that provides a paint server provides a paint source. Note: In SVG1.1, the <linearGradient>, <radialGradient>, and <pattern> elements provide paint sources. They are drawn as described in the spec, with the coordinate systems defined as follows:

The coordinate system has its origin at the top left corner of the rectangle defined by the concrete object size that it's being drawn into, and the same width and height as the concrete object size. A single user coordinate is the width and height of the concrete object size.
The coordinate system has its origin at the top left corner of the rectangle defined by the concrete object size that it's being drawn into, and the same width and height as the concrete object size. User coordinates are sized equivalently to the CSS ‘px’ unit.

It is expected that a future version of this module will define ways to refer to paint sources in external documents, or ones that are created solely by script and never inserted into a document at all.

3.4.2. Cycle Detection

The ‘element()’ function can produce nonsensical circular relationships, such as an element using itself as its own background. These relationships can be easily and reliably detected and resolved, however, by keeping track of a dependency graph and using common cycle-detection algorithms.

The dependency graph consists of edges such that:

If the graph contains a cycle, any ‘element()’ functions participating in the cycle are invalid images.

3.5. Combining images: the ‘cross-fade()’ notation

When transitioning between images, CSS requires a way to explicitly refer to the intermediate image that is a combination of the start and end images. This is accomplished with the ‘cross-fade()’ function, which indicates the two images to be combined and how far along in the transition the combination is.

Authors can also use the ‘cross-fade()’ function for many simple image manipulations, such as tinting an image with a solid color or highlighting a particular area of the page by combining an image with a radial gradient.

The syntax for ‘cross-fade()’ is defined as:

<image-combination> = cross-fade( <percentage>? <image> [, <image> | <color> ]? )

The function represents an image generated by combining two images.

The <percentage> represents how much of the first image is retained when it is blended with the second image. The <percentage> must be between ‘0%’ and ‘100%’ inclusive; any other value is invalid. If omitted, it defaults to the value ‘50%’.

If the last argument is a <color>, it represents a solid-color image with the same intrinsic dimensions as the first image. If omitted, it defaults to the color ‘transparent’.

More precisely, given ‘cross-fade(p A, B)’, where A and B are images and p is a percentage between 0% and 100%, the function represents an image with width equal to widthA × p + widthB × (1-p) and height equal to heightA × p + heightB × (1-p). The contents of the image must be constructed by first scaling A and B to the size of the generated image, then applying dissolve(A,p) plus dissolve(B,1-p). The "dissolve()" function and "plus" compositing operator are defined in the literature by Porter-Duff. [PORTERDUFF]

4. Gradients

A gradient is an image that smoothly fades from one color to another. These are commonly used for subtle shading in background images, buttons, and many other things. The gradient notations described in this section allow an author to specify such an image in a terse syntax, so that the UA can generate the image automatically when rendering the page. The syntax of a <gradient> is:

<gradient> = [ 
  <linear-gradient> | <repeating-linear-gradient> | 
  <radial-gradient> | <repeating-radial-gradient> | 
  <conic-gradient>  | <repeating-conic-gradient> ]

As with the other <image> types defined in this specification, gradients can be used in any property that accepts images. For example:

A gradient is drawn into a box with the dimensions of the concrete object size, referred to as the gradient box. However, the gradient itself has no intrinsic dimensions.

For example, if you use a gradient as a background, by default the gradient will draw into a gradient box the size of the element's padding box. If ‘background-size’ is explicitly set to a value such as ‘100px 200px’, then the gradient box will be 100px wide and 200px tall. Similarly, for a gradient used as a ‘list-style-image’, the box would be a 1em square, which is the default object size for that property.

4.1. Linear Gradients: the ‘linear-gradient()’ notation

A linear gradient is created by specifying a gradient line and then several colors placed along that line. The image is constructed by creating an infinite canvas and painting it with lines perpendicular to the gradient line, with the color of the painted line being the color of the gradient line where the two intersect. This produces a smooth fade from each color to the next, progressing in the specified direction.

4.1.1. linear-gradient() syntax

The linear gradient syntax is:

<linear-gradient> = linear-gradient(
	[ [ <angle> | to <side-or-corner> ] ,]? 
	<color-stop>[, <color-stop>]+

<side-or-corner> = [left | right] || [top | bottom]

The first argument to the function specifies the gradient line, which gives the gradient a direction and determines how color-stops are positioned. It may be omitted; if so, it defaults to ‘to bottom’.

The gradient line's direction may be specified in two ways:

using angles
For the purpose of this argument, ‘0deg’ points upward, and positive angles represent clockwise rotation, so ‘90deg’ point toward the right.
using keywords

If the argument is ‘to top’, ‘to right’, ‘to bottom’, or ‘to left’, the angle of the gradient line is ‘0deg’, ‘90deg’, ‘180deg’, or ‘270deg’, respectively.

If the argument instead specifies a corner of the box such as ‘to top left’, the gradient line must be angled such that it points into the same quadrant as the specified corner, and is perpendicular to a line intersecting the two neighboring corners of the gradient box. This causes a color-stop at 50% to intersect the two neighboring corners (see example).

Starting from the center of the gradient box, extend a line at the specified angle in both directions. The ending point is the point on the gradient line where a line drawn perpendicular to the gradient line would intersect the corner of the gradient box in the specified direction. The starting point is determined identically, but in the opposite direction.

It is expected that the next level of this module will provide the ability to define the gradient's direction relative to the current text direction and writing-mode.

[An image showing a box with a background shading gradually from white in the bottom-left corner to black in the top-right corner.  There is a line, illustrating the gradient line, angled at 45 degrees and passing through the center of the box.  The starting point and ending point of the gradient line are indicated by the intersection of the gradient line with two additional lines that pass through the bottom-left and top-right corners of the box.]

This example illustrates visually how to calculate the gradient line from the rules above. This shows the starting and ending point of the gradient line, long with the actual gradient, produced by an element with ‘background: linear-gradient(45deg, white, black);’.

Notice how, though the starting point and ending point are outside of the box, they're positioned precisely right so that the gradient is pure white exactly at the corner, and pure black exactly at the opposite corner. That's intentional, and will always be true for linear gradients.

The gradient's color stops are typically placed between the starting point and ending point on the gradient line, but this isn't required - the gradient line extends infinitely in both directions. The starting point and ending point are merely arbitrary location markers - the starting point defines where 0%, 0px, etc are located when specifying color-stops, and the ending point defines where 100% is located. Color-stops are allowed to have positions before 0% or after 100%.

The color of the gradient at any point is determined by finding the unique line passing through that point that is perpendicular to the gradient line. The point's color is the color of the gradient line at the point where this line intersects it.

4.1.2. Linear Gradient Examples

All of the following ‘linear-gradient()’ examples are presumed to be backgrounds applied to a box that is 200px wide and 100px tall.

Below are various ways of specifying a basic vertical gradient:

linear-gradient(yellow, blue);
linear-gradient(to bottom, yellow, blue);
linear-gradient(180deg, yellow, blue);
linear-gradient(to top, blue, yellow);
linear-gradient(to bottom, yellow 0%, blue 100%);

This demonstrates the use of an angle in the gradient. Note that, though the angle is not exactly the same as the angle between the corners, the gradient line is still sized so as to make the gradient yellow exactly at the upper-left corner, and blue exactly at the lower-right corner.

linear-gradient(135deg, yellow, blue);
linear-gradient(-45deg, blue, yellow);

This demonstrates a 3-color gradient, and how to specify the location of a stop explicitly:

linear-gradient(yellow, blue 20%, #0f0);

This demonstrates a corner-to-corner gradient specified with keywords. Note how the gradient is red and blue exactly in the bottom-left and top-right corners, respectively, exactly like the second example. Additionally, the angle of the gradient is automatically computed so that the color at 50% (in this case, white) stretches across the top-left and bottom-right corners.

linear-gradient(to top right, red, white, blue)

(Image requires SVG)

4.2. Radial Gradients: the ‘radial-gradient()’ notation

In a radial gradient, rather than colors smoothly fading from one side of the gradient box to the other as with linear gradients, they instead emerge from a single point and smoothly spread outward in a circular or elliptical shape.

A radial gradient is specified by indicating the center of the gradient (where the 0% ellipse will be) and the size and shape of the ending shape (the 100% ellipse). Color stops are given as a list, just as for ‘linear-gradient()’. Starting from the center and progressing towards (and potentially beyond) the ending shape uniformly-scaled concentric ellipses are drawn and colored according to the specified color stops.

4.2.1. radial-gradient() Syntax

The radial gradient syntax is:

<radial-gradient> = radial-gradient(
  [ [ <shape> || <size> ] [ at <position> ]? , |
    at <position>, 
  <color-stop> [ , <color-stop> ]+

Here is an example of a circular radial gradient 5em wide and positioned with its center in the top left corner:

radial-gradient(5em circle at top left, yellow, blue)

We should add the ability to move the focus of the gradient, as in the original -webkit-gradient() function. See proposal in http://lists.w3.org/Archives/Public/www-style/2011Nov/0210.html for "from <position>" and "from offset <offset>".

The arguments are defined as follows:

Determines the center of the gradient. The <position> value type (which is also used for ‘background-position’) is defined in [CSS3VAL], and is resolved using the center-point as the object area and the gradient box as the positioning area. If this argument is omitted, it defaults to ‘center’.
Can be either ‘circle’ or ‘ellipse’; determines whether the gradient's ending shape is a circle or an ellipse, respectively. If <shape> is omitted, the ending shape defaults to a circle if the <size> is a single <length>, and to an ellipse otherwise.

Determines the size of the gradient's ending shape. If omitted it defaults to ‘farthest-corner’. It can be given explicitly or by keyword. For the purpose of the keyword definitions, consider the gradient box edges as extending infinitely in both directions, rather than being finite line segments.

If the ending-shape is an ellipse, its axises are aligned with the horizontal and vertical axises.

Both ‘circle’ and ‘ellipse’ gradients accept the following keywords as their <size>:

The ending shape is sized so that it exactly meets the side of the gradient box closest to the gradient's center. If the shape is an ellipse, it exactly meets the closest side in each dimension.
Same as ‘closest-side’, except the ending shape is sized based on the farthest side(s).
The ending shape is sized so that it passes through the corner of the gradient box closest to the gradient's center. If the shape is an ellipse, the ending shape is given the same aspect-ratio it would have if ‘closest-side’ were specified.
Same as ‘closest-corner’, except the ending shape is sized based on the farthest corner. If the shape is an ellipse, the ending shape is given the same aspect ratio it would have if ‘farthest-side’ were specified.

If <shape> is specified as ‘circle’ or is omitted, the <size> may be given explicitly as:


Gives the radius of the circle explicitly. Negative values are invalid.

Note that percentages are not allowed here; they can only be used to specify the size of an elliptical gradient, not a circular one. This restriction exists because there is are multiple reasonable answers as to which dimension the percentage should be relative to. A future level of this module may provide the ability to size circles with percentages, perhaps with more explicit controls over which dimension is used.

If <shape> is specified as ‘ellipse’ or is omitted, <size> may instead be given explicitly as:

[ <length> | <percentage> ]{2}
Gives the size of the ellipse explicitly. The first value represents the horizontal radius, the second the vertical radius. Percentages values are relative to the corresponding dimension of the gradient box. Negative values are invalid.

Expanded with the above definitions, the grammar becomes:

<radial-gradient> = radial-gradient(
  [ [ circle               || <length> ]                          [ at <position> ]? , | 
    [ ellipse              || [ <length> | <percentage> ]{2} ]    [ at <position> ]? , |
    [ [ circle | ellipse ] || <extent-keyword> ]                  [ at <position> ]? , |
    at <position> ,
  <color-stop> [ , <color-stop> ]+
<extent-keyword> = closest-corner | closest-side | farthest-corner | farthest-side

4.2.2. Placing Color Stops

Color-stops are placed on a gradient ray, similar to the gradient line of linear gradients. The gradient ray is anchored at the center of the gradient and extends toward the right. The 0% location is at the start of the gradient ray, and the 100% location is on the point where the gradient ray intersects the ending shape. A color-stop can be placed at a negative location; though the negative region of the gradient ray is never directly consulted for rendering, color stops placed there can affect the color of non-negative locations on the gradient ray through interpolation or repetition (see repeating gradients). For example, ‘radial-gradient(red -50px, yellow 100px)’ produces an elliptical gradient that starts with a reddish-orange color in the center (specifically, #f50) and transitions to yellow. Locations greater than 100% simply specify a location a correspondingly greater distance from the center of the gradient.

The color of the gradient at any point is determined by first finding the unique ellipse passing through that point with the same center, orientation, and ratio between major and minor axises as the ending-shape. The point's color is then the color of the gradient ray at the location where this ellipse intersects it.

4.2.3. Degenerate Radial Gradients

Some combinations of position, size, and shape will produce a circle or ellipse with a radius of 0. This will occur, for example, if the center is on a gradient box edge and ‘closest-side’ or ‘closest-corner’ is specified or if the size and shape are given explicitly and either of the radiuses is zero. In these degenerate cases, the gradient must be be rendered as follows:

If the ending shape is a circle with zero radius:
Render as if the ending shape was a circle whose radius was an arbitrary very small number greater than zero. This will make the gradient continue to look like a circle.
If the ending shape has zero width (regardless of the height):
Render as if the ending shape was an ellipse whose height was an arbitrary very large number and whose width was an arbitrary very small number greater than zero. This will make the gradient look similar to a horizontal linear gradient that is mirrored across the center of the ellipse. It also means that all color-stop positions specified with a percentage resolve to ‘0px’.
Otherwise, if the ending shape has zero height:
Render as if the ending shape was an ellipse whose width was an arbitrary very large number and whose height was an arbitrary very small number greater than zero. This will make the gradient look like a solid-color image equal to the color of the last color-stop, or equal to the average color of the gradient if it's repeating.

4.2.4. Radial Gradient Examples

All of the following examples are applied to a box that is 200px wide and 100px tall.

These examples demonstrate different ways to write the basic syntax for radial gradients:

radial-gradient(yellow, green);
radial-gradient(ellipse at center, yellow 0%, green 100%);
radial-gradient(farthest-corner at 50% 50%, yellow, green);

radial-gradient(circle, yellow, green);

radial-gradient(red, yellow, green);

This image shows a gradient originating from somewhere other than the center of the box:

radial-gradient(farthest-side at left bottom, red, yellow 50px, green);

Here we illustrate a ‘closest-side’ gradient.

radial-gradient(closest-side at 20px 30px, red, yellow, green);
radial-gradient(20px 30px at 20px 30px, red, yellow, green);

radial-gradient(closest-side circle at 20px 30px, red, yellow, green);
radial-gradient(20px 20px at 20px 30px, red, yellow, green);

4.3. Conic Gradients: the ‘conic-gradient()’ notation

A conic gradient starts by specifying the center of a circle, similar to radial gradients, except that conic gradient color-stops are placed around the circumference of the circle, rather than on a line emerging from the center, causing the color to smoothly transition as you spin around the center, rather than as you progress outward from the center.

A conic gradient is specified by indicating the center of the gradient, and then specifying a list of color-stops. Unlike linear and radial gradients, whose color-stops are placed by specifying a <length>, the color-stops of a conic gradient are specified with an <angle>. Rays are then drawn emerging from the center and pointing in all directions, with the color of each ray equal to the color of the gradient-line where they intersect it.

These gradients are called "conic" or "conical" because, if the color stops are chosen to be significantly lighter on one side than the other, it produces a pattern that looks like a cone observed from above.

4.3.1. conic-gradient()’ Syntax

The syntax for a conic gradient is:

<conic-gradient> = conic-gradient(
  [ at <position> , ]?
  <angular-color-stop> [ , <angular-color-stop> ]+

The <position> argument is defined in [!CSS3VAL], and is resolved using the center-point as the object area and the gradient box as the positioning area. If this argument is omitted, it defaults to ‘at center’.

The <angular-color-stop> is defined in the section on color-stops.

Anything else that might be useful? Defining the shape of the gradient as elliptical, perhaps?

4.3.2. Placing Color Stops

Color stops are placed on a gradient line that curves around the center in a circle, with both the 0% and 100% locations at 0deg. Just like linear gradients, 0deg points to the top of the page, and increasing angles correspond to clockwise movement around the circle.

It may be more helpful to think of the gradient line as forming a spiral, where only the segment from 0deg to 360deg is rendered. This avoids any confusion about "overlap" when you have angles outside of the rendered region.

A color-stop can be placed at a location before 0% or after 100%; though these regions are never directly consulted for rendering, color stops placed there can affect the color of color-stops within the rendered region through interpolation or repetition (see repeating gradients). For example, ‘conic-gradient(red -50%, yellow 150%)’ produces a conic gradient that starts with a reddish-orange color at 0deg (specifically, #f50), and transitions to an orangish-yellow color at 360deg (specifically, #fa0).

The color of the gradient at any point is determined by first finding the unique ray anchored at the center of the gradient that passes through the given point. The point's color is then the color of the gradient line at the location where this ray intersects it.

4.3.3. Conic Gradient Examples

Produce examples. Better yet, strike this section, and intermix some examples into the sections above. Do this for the other two types of gradients as well.

4.4. Repeating Gradients: the ‘repeating-linear-gradient()’, ‘repeating-radial-gradient()’, and ‘repeating-conic-gradient()’ notations

In addition to <linear-gradient>, <radial-gradient>, and <conic-gradient>, this specification defines <repeating-linear-gradient>, <repeating-radial-gradient>, and <repeating-conic-gradient> values. These notations take the same values and are interpreted the same as their respective non-repeating siblings defined previously.

When rendered, however, the color-stops are repeated infinitely in both directions, with their positions shifted by multiples of the difference between the last specified color-stop's position and the first specified color-stop's position. For example, ‘repeating-linear-gradient(red 10px, blue 50px)’ is equivalent to ‘linear-gradient(..., red -30px, blue 10px, red 10px, blue 50px, red 50px, blue 90px, ...)’. Note that the last color-stop and first color-stop will always coincide at the boundaries of each group, which will produce sharp transitions if the gradient does not start and end with the same color.

Repeating gradient syntax is identical to that of non-repeating gradients:

repeating-linear-gradient(red, blue 20px, red 40px)

repeating-radial-gradient(red, blue 20px, red 40px)

repeating-radial-gradient(circle closest-side at 20px 30px, red, yellow, green 100%, yellow 150%, red 200%)

repeating-conic-gradient(at 20%, white 0deg, white 20deg, red 20deg, red 40deg)

Insert rendering here.

If the distance between the first and last color-stops is non-zero, but is small enough that the implementation knows that the physical resolution of the output device is insufficient to faithfully render the gradient, the implementation must find the average color of the gradient and render the gradient as a solid-color image equal to the average color.

If the distance between the first and last color-stops is zero (or rounds to zero due to implementation limitations), the implementation must find the average color of a gradient with the same number and color of color-stops, but with the first and last color-stop an arbitrary non-zero distance apart, and the remaining color-stops equally spaced between them. Then it must render the gradient as a solid-color image equal to that average color.

If the height of a repeating radial gradient is zero, or is close enough to zero that the implementation knows that the physical resolution of the output device is insufficient to faithfully render the gradient, the implementation must find the average color of the gradient and render the gradient as a solid-color image equal to the average color.

To find the average color of a gradient, run these steps:

  1. Define list as an initially-empty list of premultiplied RGBA colors, and total-length as the distance between first and last color stops.
  2. For each adjacent pair of color-stops, define weight as half the distance between the two color-stops, divided by total-length. Add two entries to list, the first obtained by representing the color of the first color-stop in premultiplied sRGBA and scaling all of the components by weight, and the second obtained in the same way with the second color-stop.
  3. Sum the entries of list component-wise to produce the average color, and return it.

As usual, implementations may use whatever algorithm they wish, so long as it produces the same result as the above.

For example, the following gradient is rendered as a solid light-purple image (equal to rgb(75%,50%,75%)):

repeating-linear-gradient(red 0px, white 0px, blue 0px);

The following gradient would render the same as the previous under normal circumstances (because desktop monitors can't faithfully render color-stops 1/10th of a pixel apart), but would render as a normal repeating gradient if, for example, the author applied "zoom:100;" to the element on which the gradient appears:

repeating-linear-gradient(red 0px, white .1px, blue .2px);

4.5. Gradient Color-Stops

<color-stop> = <color> [ <percentage> | <length> ]{0,2}
<angular-color-stop> = <color> [ <percentage> | <angle> ]{0,2}

Color-stops are points placed along the gradient line defined by the gradient function in which the color-stop is used. Color-stops must be specified in order. Percentages refer to the length of the gradient line, with 0% being at the starting point and 100% being at the ending point. Lengths are measured from the starting point in the direction of the ending point. Angles are measured with 0deg pointing up, and positive angles corresponding to clockwise rotations. Color-stops are usually placed between the starting point and ending point, but that's not required; the gradient line extends infinitely in both directions, and a color-stop can be placed at any position on the line.

A color-stop with two locations is exactly equivalent to specifying two color-stops with the same color, one at each location, including how it reacts to the clean-up steps below. Specifying two locations makes it easier to create solid-color "stripes" in a gradient, without having to repeat the color twice.

Make sure this is handled well in the serialization rules.

At each color-stop, the line is the color of the color-stop. Between two color-stops, the line's color is linearly interpolated between the colors of the two color-stops, with the interpolation taking place in premultiplied RGBA space. Before the first color-stop, the line is the color of the first color-stop. After the last color-stop, the line is the color of the last color-stop.

The following steps must be applied in order to process the list of color-stops. After applying these rules, all color-stops will have a definite position and they will be in ascending order:

  1. If the first color-stop does not have a position, set its position to 0%. If the last color-stop does not have a position, set its position to 100%.
  2. If a color-stop has a position that is less than the specified position of any color-stop before it in the list, set its position to be equal to the largest specified position of any color-stop before it.
  3. If any color-stop still does not have a position, then, for each run of adjacent color-stops without positions, set their positions so that they are evenly spaced between the preceding and following color-stops with positions.

This requires us to wait until after layout to do fix-up, because implied-position stops (set by step 3) may depend on stops that need layout information to place, and which may be corrected by step 2. Swapping steps 2 and 3 would let us interpolate color-stops purely at computed-value time, which is a nice plus, at the cost of changing behavior from level 3 for some edge cases that triggered fixup.

If multiple color-stops have the same position, they produce an infinitesimal transition from the one specified first in the rule to the one specified last. In effect, the color suddenly changes at that position rather than smoothly transitioning.

Below are several pairs of gradients. The latter of each pair is a manually "fixed-up" version of the former, obtained by applying the above rules. For each pair, both gradients will render identically. The numbers in each arrow specify which fixup steps are invoked in the transformation.

1. linear-gradient(red, white 20%, blue)
   linear-gradient(red 0%, white 20%, blue 100%)

2. linear-gradient(red 40%, white, black, blue)
   linear-gradient(red 40%, white 60%, black 80%, blue 100%)

3. linear-gradient(red -50%, white, blue)
   linear-gradient(red -50%, white 25%, blue 100%)

4. linear-gradient(red -50px, white, blue)
   linear-gradient(red -50px, white calc(-25px + 50%), blue 100%)

5. linear-gradient(red 20px, white 0px, blue 40px)
   linear-gradient(red 20px, white 20px, blue 40px)

6. linear-gradient(red, white -50%, black 150%, blue)
   linear-gradient(red 0%, white 0%, black 150%, blue 150%)

7. linear-gradient(red 80px, white 0px, black, blue 100px)
   linear-gradient(red 80px, white 80px, black 90px, blue 100px)

The following example illustrates the difference between a gradient transitioning in pre-multiplied sRGBA and one transitioning (incorrectly) in non-premultiplied. In both of these example, the gradient is drawn over a white background. Both gradients could be written with the following value:

linear-gradient(90deg, red, transparent, blue)

In premultiplied space, transitions to or from "transparent" always look nice:

(Image requires SVG)

On the other hand, if a gradient were to incorrectly transition in non-premultiplied space, the colors near "transparent" would noticeably darken to a grayish color, because "transparent" is actually a shorthand for ‘rgba(0,0,0,0)’, or transparent black:

(Image requires SVG)

Note: It is recommended that authors not mix different types of units, such as px, em, or %, in a single rule, as this can cause a color-stop to unintentionally try to move before an earlier one. For example, the rule ‘background-image: linear-gradient(yellow 100px, blue 50%)’ wouldn't require any fix-up as long as the background area is at least 200px tall. If it was 150px tall, however, the blue color-stop's position would be equivalent to "75px", which precedes the yellow color-stop, and would be corrected to a position of 100px.

Note: The definition and implications of "premultiplied" color spaces are given elsewhere in the technical literature, but a quick primer is given here to illuminate the process. Given a color expressed as an rgba() 4-tuple, one can convert this to a premultiplied representation by multiplying the red, green, and blue components by the alpha component. For example, a partially-transparent blue may be given as rgba(0,0,255,.5), which would then be expressed as [0, 0, 127.5, .5] in its premultiplied representation. Interpolating colors using the premultiplied representations rather than the plain rgba representations tends to produce more attractive transitions, particularly when transitioning from a fully opaque color to fully transparent. Note that transitions where either the transparency or the color are held constant (for example, transitioning between rgba(255,0,0,100%) and rgba(0,0,255,100%) or rgba(255,0,0,100%) and rgba(255,0,0,0%)) have identical results whether the color interpolation is done in premultiplied or non-premultiplied color-space. Differences only arise when both the color and transparency differ between the two endpoints.

5. Sizing Images and Objects in CSS

Images used in CSS may come from a number of sources: from binary image formats (such as gif, jpeg, etc), dedicated markup formats (such as SVG), and CSS-specific formats (such as the linear-gradient() value type defined in this specification). As well, a document may contain many other types of objects, such as video, plugins, or nested documents. These images and objects (just objects hereafter) may offer many types of sizing information to CSS, or none at all. This section defines generically the size negotiation model between the object and the CSS layout algorithms.

5.1. Object-Sizing Terminology

In order to define this handling, we define a few terms, to make it easier to refer to various concepts:

intrinsic dimensions

The term intrinsic dimensions refers to the set of the intrinsic height, intrinsic width, and intrinsic aspect ratio (the ratio between the width and height), each of which may or may not exist for a given object. These intrinsic dimensions represent a preferred or natural size of the object itself; that is, they are not a function of the context in which the object is used. CSS does not define how the intrinsic dimensions are found in general.

Raster images are an example of an object with all three intrinsic dimensions. SVG images designed to scale might have only an intrinsic aspect ratio; SVG images can also be created with only an intrinsic width or height. CSS gradients, defined in this specification, are an example of an object with no intrinsic dimensions at all. Another example of this is embedded documents, such as the <iframe> element in HTML. An object cannot have only two intrinsic dimensions, as any two automatically define the third.

If an object (such as an icon) has multiple sizes, then the largest size (by area) is taken as its intrinsic size. If it has multiple aspect ratios at that size, or has multiple aspect ratios and no size, then the aspect ratio closest to the aspect ratio of the default object size is used. Determine this by seeing which aspect ratio produces the largest area when fitting it within the default object size using a contain fit; if multiple sizes tie for the largest area, the wider size is chosen as its intrinsic size.

specified size
The specified size of an object is given by CSS, such as through the ‘width’ and ‘height’ or ‘background-size’ properties. The specified size can be a definite width and height, a set of constraints, or a combination thereof.
concrete object size
The concrete object size is the result of combining an object's intrinsic dimensions and specified size with the default object size of the context it's used in, producing a rectangle with a definite width and height.
default object size

The default object size is a rectangle with a definite height and width used to determine the concrete object size when both the intrinsic dimensions and specified size are missing dimensions.

5.2. CSS⇋Object Negotiation

Objects in CSS are sized and rendered as follows:

  1. When an image or object is specified in a document, such as through a ‘url()’ value in a ‘background-image’ property or a src attribute on an <img> element, CSS queries the object for its intrinsic dimensions.
  2. Using the intrinsic dimensions, the specified size, and the default object size for the context the image or object is used in, CSS then computes a concrete object size. (See the following section.) This defines the size and position of the region the object will render in.
  3. CSS asks the object to render itself at the concrete object size. CSS does not define how objects render when the concrete object size is different from the object's intrinsic dimensions. The object may adjust itself to match the concrete object size in some way, or even render itself larger or smaller than the concrete object size to satisfy sizing constraints of its own.
  4. Unless otherwise specified by CSS, the object is then clipped to the concrete object size.

5.3. Concrete Object Size Resolution

Currently the rules for sizing objects are described in each context that such objects are used. This section defines some common sizing constraints and how to resolve them so that future specs can refer to them instead of redefining size resolution in each instance.

5.3.1. Default Sizing Algorithm

The default sizing algorithm is a set of rules commonly used to find an object's concrete object size. It resolves the simultaneous constraints presented by the object's intrinsic dimensions and either an unconstrained specified size or one consisting of only a definite width and/or height.

Some object sizing rules (such as those for ‘list-style-image’) correspond exactly to the default sizing algorithm. Others (such as those for ‘border-image’) invoke the default sizing algorithm but also apply additional sizing rules before arriving at a final concrete object size.

The default sizing algorithm is defined as follows:

5.3.2. Cover and Contain Constraint Sizing

Two other common specified sizes are the contain constraint and the cover constraint, both of which are resolved against a specified constraint rectangle using the object's intrinsic aspect ratio:

In both cases, if the object doesn't have an intrinsic aspect ratio, the concrete object size is the specified constraint rectangle.

5.4. Examples of CSS Object Sizing

The following examples show how the CSS 2.1 and CSS3 Backgrounds & Borders sizing algorithms correspond to concepts defined in this specification.

The rules for calculating the concrete object size of a background are defined in CSS2.1§14.2.1 and CSS3BG§3.9. CSS2.1 uses the default sizing algorithm with no specified size and the background positioning area as the default object size. [CSS21] In CSS3, ‘background-size’ property can give a sizing constraint, invoking either the default sizing algorithm or one of the contain or cover constraints. The concrete object size is further adjusted in later steps if ‘background-repeat’ has a ‘round’ value. [CSS3BG]
The rules for calculating the concrete object size of a list-style image are defined in CSS2.1§12.5.1. They use the default sizing algorithm with no specified size and a default object size of 1em square.
Border images are sized twice: first the entire image is sized to determine the slice points, then the slices are sized to decorate the border. The first sizing operation is defined in CSS3BG§6.2 and uses the default sizing algorithm with no specified size, and the border image area as the default object size. The second operation is defined in CSS3BG§6.2: the default sizing algorithm is used to determine an initial size for each slice with the corresponding border image area part as the default object size. By default the specified size matches this default object size; however the ‘border-image-repeat’ property can drop the specified size in one or more directions and may also apply an additional rounding step. [CSS3BG]
The rules for calculating the concrete object size of a cursor are defined in CSS2.1 § 18.1: Cursors. The default object size is a UA-defined size that should be based on the size of a typical cursor on the UA's operating system. [CSS21]
Objects inserted via the CSS2.1 ‘content’ property are anonymous replaced elements, and are sized the same way. [CSS21] Note that such anonymous elements have all their non-inherited properties (including ‘width’, ‘height’, etc.) set to their initial values.
replaced elements
CSS 2.1 defines the sizing of replaced elements (including those inserted as generated content via ‘content’) in sections 10.3.2, 10.4, 10.6.2, and 10.7. [CSS21] The ‘object-fit’ property defined below defines how the concrete object size corresponds to the element's used width and height; by default they coincide.

5.5. Sizing Objects: the ‘object-fit’ property

Name: object-fit
Value: fill | contain | cover | none | scale-down
Initial: fill
Applies to: replaced elements
Inherited: no
Percentages: N/A
Media: visual
Computed value: specified value
Animatable: no

The ‘object-fit’ property specifies how the contents of a replaced element should be fitted to the box established by its used height and width.

The replaced content is sized to fill the element's content box: the object's concrete object size is the element's used width and height.
The replaced content is sized to maintain its aspect ratio while fitting within the element's content box: its concrete object size is resolved as a contain constraint against the element's used width and height.
The replaced content is sized to maintain its aspect ratio while filling the element's entire content box: its concrete object size is resolved as a cover constraint against the element's used width and height.
The replaced content is not resized to fit inside the element's content box: determine the object's concrete object size using the default sizing algorithm with no specified size, and a default object size equal to the replaced element's used width and height.

Size the content as if ‘none’ or ‘contain’ were specified, whichever would result in a smaller concrete object size.

Note that both ‘none’ and ‘contain’ respect the content's intrinsic aspect ratio, so the concept of "smaller" is well-defined.

If the content does not completely fill the replaced element's content box, the unfilled space shows the replaced element's background. Since replaced elements always clip their contents to the content box, the content will never overflow. See the ‘object-position’ property for positioning the object with respect to the content box.

An example showing how four of the values of ‘object-fit’ cause the replaced element (blue figure) to be scaled to fit its height/width box (shown with a green background), using the initial value for ‘object-position’. The fifth value, ‘scale-down’, in this case looks identical to ‘contain’.

Note: the ‘object-fit’ property has similar semantics to the fit attribute in [SMIL10] and the <meetOrSlice> parameter on the preserveAspectRatio attribute in [SVG11].

Note: Per the CSS⇋Object Negotiation algorithm, the concrete object size (or, in this case, the size of the content) does not directly scale the object itself - it is merely passed to the object as information about the size of the visible canvas. How to then draw into that size is up to the image format. In particular, raster images always scale to the given size, while SVG uses the given size as the size of the "SVG Viewport" (a term defined by SVG) and then uses the values of several attributes on the root <svg> element to determine how to draw itself.

5.6. Positioning Objects: the ‘object-position’ property

Name: object-position
Value: <position>
Initial: 50% 50%
Applies to: replaced elements
Inherited: no
Percentages: refer to width and height of box itself
Media: visual
Computed value: specified value
Animatable: yes
Canonical Order: the horizontal component of the <position>, followed by the vertical component

The ‘object-position’ property determines the alignment of the replaced element inside its box. The <position> value type (which is also used for ‘background-position’) is defined in [CSS3VAL], and is resolved using the concrete object size as the object area and the content box as the positioning area.

Note that areas of the box not covered by the replaced element will show the element's background.

6. Image Processing

6.1. Overriding Image Resolutions: the ‘image-resolution’ property

The image resolution is defined as the number of image pixels per unit length, e.g., pixels per inch. Some image formats can record information about the resolution of images. This information can be helpful when determining the actual size of the image in the formatting process. However, the information can also be wrong, in which case it should be ignored. By default, CSS assumes a resolution of one image pixel per CSS ‘px’ unit; however, the ‘image-resolution’ property allows using some other resolution.

Name: image-resolution
Value: [from-image || <resolution>] && snap?
Initial: 1dppx
Applies to: all elements
Inherited: yes
Media: visual
Computed value: as specified, except with ‘<resolution>’ possibly altered by computing for ‘snap’ (see below)
Animatable: no
Canonical Order: per grammar

The ‘image-set()’ notation can alter the intrinsic resolution of an image, which ideally would be automatically honored without having to set this property. How should we best address this? Change the initial value to ‘auto’, meaning "1dppx, unless CSS says otherwise"? Say that image-resolution has no effect on images whose resolution was set by something else in CSS? Or somehow wordsmithing ‘image-set()’ in some way such that it always produces ‘1dppx’ images somehow?

The ‘image-resolution’ property specifies the intrinsic resolution of all raster images used in or on the element. It affects both content images (e.g. replaced elements and generated content) and decorative images (such as ‘background-image’). The intrinsic resolution of an image is used to determine the image's intrinsic dimensions. Values have the following meanings:

Specifies the intrinsic resolution explicitly. A "dot" in this case corresponds to a single image pixel.
The image's intrinsic resolution is taken as that specified by the image format. If the image does not specify its own resolution, the explicitly specified resolution is used (if given), else it defaults to ‘1dppx’.
If the "snap" keyword is provided, the computed ‘<resolution>’ (if any) is the specified resolution rounded to the nearest value that would map one image pixel to an integer number of device pixels. If the resolution is taken from the image, then the used intrinsic resolution is the image's native resolution similarly adjusted.

As vector formats such as SVG do not have an intrinsic resolution, this property has no effect on vector images.

Printers tend to have substantially higher resolution than computer monitors; due to this, an image that looks fine on the screen may look pixellated when printed out. The ‘image-resolution’ property can be used to embed a high-resolution image into the document and maintain an appropriate size, ensuring attractive display both on screen and on paper:

img.high-res {
	image-resolution: 300dpi;

With this set, an image meant to be 5 inches wide at 300dpi will actually display as 5in wide; without this set, the image would display as approximately 15.6in wide since the image is 15000 image pixels across, and by default CSS displays 96 image pixels per inch.

Some image formats can encode the image resolution into the image data. This rule specifies that the UA should use the image resolution found in the image itself, falling back to 1 image pixel per CSS ‘px’ unit.

img { image-resolution: from-image }

These rules both specify that the UA should use the image resolution found in the image itself, but if the image has no resolution, the resolution is set to ‘300dpi’ instead of the default ‘1dppx’.

img { image-resolution: from-image 300dpi }
img { image-resolution: 300dpi from-image }

Using this rule, the image resolution is set to 300dpi. (The resolution in the image, if any, is ignored.)

img { image-resolution: 300dpi }

This rule, on the other hand, if used when the screen's resolution is 96dpi, would instead render the image at 288dpi (so that 3 image pixels map to 1 device pixel):

img { image-resolution: 300dpi snap; }

The ‘snap’ keyword can also be used when the resolution is taken from the image:

img { image-resolution: snap from-image; }

An image declaring itself as 300dpi will, in the situation above, display at 288dpi (3 image pixels per device pixel) whereas an image declaring 72dpi will render at 96dpi (1 image pixel per device pixel).

6.2. Orienting an Image on the Page: the ‘image-orientation’ property

If a picture is taken with a camera turned on its side, or a document isn't positioned correctly within a scanner, the resultant image may be "sideways" or even upside-down. The ‘image-orientation’ property provides a way to apply an "out-of-band" rotation to image source data to correctly orient an image.

Note that this property is not intended to specify layout transformations such as arbitrary rotation or flipping the image in the horizontal or vertical direction. (See [CSS3-2D-TRANSFORMS] for a feature designed to do that.) It is also not needed to correctly orient an image when printing in landscape versus portrait orientation, as that rotation is done as part of layout. (See [CSS3PAGE].) It should only be used to correct incorrectly-oriented images.

Name: image-orientation
Value: <angle> | from-image
Initial: 0deg
Applies to: all elements
Inherited: yes
Media: visual
Computed value: an <angle>, rounded and normalized (see text)
Animatable: no

This property specifies an orthogonal rotation to be applied to an image before it is laid out. It applies only to content images (e.g. replaced elements and generated content), not decorative images (such as ‘background-image’). CSS layout processing applies to the image after rotation. This implies, for example:

Values have the following meanings:

Positive values cause the image to be rotated to the right (in a clockwise direction), while negative values cause a rotation to the left. The computed value of the property is calculated by rounding the specified angle to the nearest quarter-turn (90deg, 100grad, .25turn, etc.), rounding away from 0 (that is, 45deg is rounded to 90deg, while -45deg is rounded to -90deg), then moduloing the value by 1 turn (360deg, 400grad, etc.).
If the image has an orientation specified in its metadata, such as EXIF, this value computes to the angle that the metadata specifies is necessary to correctly orient the image. If necessary, this angle is then rounded and normalized as described above for an <angle> value. If there is no orientation specified in its metadata, this value computes to ‘0deg’.

Do we need to handle the "unnatural" EXIF orientations? These might be used by, say, front-facing cameras (that take a picture of the user and therefore display on the screen mirrored). This would imply adding a ‘flip’ keyword to go along with <angle>.

The ‘image-orientation’ property must be applied before any other transformations, such as being specified in the ‘image()’ function with an opposite directionality to its context, or using CSS Transforms.

The following example rotates the image 90 degrees clockwise:

img.ninety     { image-orientation: 90deg }
<img class="ninety" src=... />

The same effect could be achieved with, for example, an angle of -270deg or 450deg.

6.3. Determing How To Scale an Image: the ‘image-rendering’ property

Name: image-rendering
Value: auto | crisp-edges | pixelated
Initial: auto
Applies to: all elements
Inherited: yes
Media: visual
Computed value: specified value
Animatable: no

The ‘image-rendering’ property provides a hint to the user-agent about what aspects of an image are most important to preserve when the image is scaled, to aid the user-agent in the choice of an appropriate scaling algorithm. When specified on an element, it applies to all images given in properties for the element, such as background images, list-style images, or the content of replaced elements when they represent an image that must be scaled. The values of the ‘image-rendering’ property are interpreted as follows:

The image should be scaled with an algorithm that maximizes the appearance of the image. In particular, scaling algorithms that "smooth" colors are acceptable, such as bilinear interpolation. This is intended for images such as photos.
The image must be scaled with an algorithm that preserves contrast and edges in the image, and which does not smooth colors or introduce blur to the image in the process. This is intended for images such as pixel art.
When scaling the image up, the "nearest neighbor" or similar algorithm must be used, so that the image appears to be simply composed of very large pixels. When scaling down, this is the same as auto.

Should we move the current ‘auto’ behavior to a well-named keyword (‘smooth’?) and just define ‘auto’ to be "whatever the UA wants", which is what ‘auto’ usually means?

This property does not dictate any particular scaling algorithm to be used. For example, with ‘image-rendering:auto’, a user agent might scale images with bilinear interpolation by default, switch to nearest-neighbor interpolation in high-load situations, and switch to a high-quality scaling algorithm like Lanczos interpolation for static images that aren't moving or changing. Similarly, with ‘image-rendering:crisp-edges’, a user agent might scale images with nearest-neighbor interpolation by default, and switch to EPX interpolation in low-load situations.

This property previously accepted the values ‘optimizeSpeed’ and ‘optimizeQuality’. These are now deprecated; a user agent must accept them as valid values but must treat them as having the same behavior as auto, and authors must not use them.

7. Interpolation

This section describes how to interpolate between new value types defined in this specification, for use with modules such as CSS Transitions and CSS Animations.

If an algorithm below simply states that two values should be "interpolated" or "transitioned" without further details, then the value should be interpolated as described by the Transitions spec. Otherwise, the algorithm may reference a variable "t" in its detailed description of the interpolation. This is a number which starts at 0% and goes to 100%, and is set to a value that represents the progress through the transition, based on the duration of the transition, the elapsed time, and the timing function in use. For example, with a linear timing function and a 1s duration, after .3s t is equal to 30%.

7.1. Interpolating <image>

All images can be interpolated, though some special types of images (like some gradients) have their own special interpolation rules. In general terms, images are interpolated by scaling them to the size of the start image and cross-fading the two while they transition to the size of the end image.

In specific terms, at each point in the interpolation the image is equal to cross-fade( (100% - t) <start image>, <end image>).

Special-case interpolating to/from no image, like "background-image: url(foo);" to "background-image: none;".

7.2. Interpolating <image-combination>

If both the starting and ending images are <image-combination>s which differ only by by their <percentage>, they must be interpolated by interpolating their <percentage>. Otherwise, they must be interpolated as generic <image>s.

7.3. Interpolating <gradient>

This section needs review and improvement. In particular, I believe the handling of linear-gradient() is incomplete - I think we want to specifically interpolate the "length" of the gradient line (the distance between 0% and 100%) between the starting and ending positions explicitly, so it doesn't grow and then shrink over a single animation.

Gradient images can be interpolated directly in CSS transitions and animations, smoothly animating from one gradient to another. There are only a few restrictions on what gradients are allowed to be interpolated:

  1. Both the starting and ending gradient must be expressed with the same function. (For example, you can transition from a ‘linear-gradient()’ to a ‘linear-gradient()’, but not from a ‘linear-gradient()’ to a ‘radial-gradient()’ or a ‘repeating-linear-gradient()’.)
  2. Both the starting and ending gradient must have the same number of <color-stop>s. For this purpose, a repeating gradient is considered to have infinite <color-stop>s, and thus all repeating gradients have the same number of <color-stop>s. (Note that one may pad a gradient with additional <color-stop>s placed atop each other, if necessary to make two gradients have the same number of <color-stop>s.)

If the two gradients satisfy both of those constraints, they must be interpolated as described below. If not, they must be interpolated as a generic <image>.

  1. Convert both the start and end gradients to their explicit forms:
    For linear gradients:
    • If the direction is specified as an <angle>, it is already in its explicit form.
    • Otherwise, change its direction to an <angle> in [‘0deg’,‘360deg’) that would produce an equivalent rendering.

      If both the start and end gradients had their direction specified with keywords, and the absolute difference between the angles their directions mapped to is greater than 180deg, add 360deg to the direction of the gradient with the smaller angle. This ensures that a transition from, for example, "to left" (270deg) to "to top" (0deg) rotates the gradient a quarter-turn clockwise, as expected, rather than rotating three-quarters of a turn counter-clockwise.

    For radial gradients:
    • If the size is specified as two <length>s or <percentage>s, it is already in its explicit form.
    • Otherwise, the size must be changed to a pair of <length>s that would produce an equivalent ending-shape. If the <shape> was specified as ‘circle’, change it to ‘ellipse’.
  2. Interpolate each component and color-stop of the gradients independently. For linear gradients, the only component is the angle. For radial gradients, the components are the horizontal and vertical position of the center and the horizontal and vertical axis lengths.
  3. To interpolate a color-stop, first match each color-stop in the start gradient to the corresponding color-stop at the same index in the end gradient. For repeating gradients, the first specified color-stop in the start and end gradients are considered to be at the same index, and all other color-stops following and preceding are indexed appropriately. Then, for each pair of color-stops, interpolate the position and color independently.

8. Serialization

This section describes the serialization of all new properties and value types introduced in this specification, for the purpose of interfacing with the CSS Object Model [CSSOM].

To serialize any function defined in this module, serialize it per its individual grammar, in the order its grammar is written in, omitting components when possible without changing the meaning, joining space-separated tokens with a single space, and following each serialized comma with a single space.

For ‘cross-fade()’, always serialize the <percentage>.

For example, a gradient specified as:

Linear-Gradient( to bottom, red 0%,yellow,black 100px)

...must serialize as:

linear-gradient(red, yellow, black 100px)

The serialization of <color> is still undefined.

9. Conformance

9.1. Document Conventions

Conformance requirements are expressed with a combination of descriptive assertions and RFC 2119 terminology. The key words “MUST”, “MUST NOT”, “REQUIRED”, “SHALL”, “SHALL NOT”, “SHOULD”, “SHOULD NOT”, “RECOMMENDED”, “MAY”, and “OPTIONAL” in the normative parts of this document are to be interpreted as described in RFC 2119. However, for readability, these words do not appear in all uppercase letters in this specification.

All of the text of this specification is normative except sections explicitly marked as non-normative, examples, and notes. [RFC2119]

Examples in this specification are introduced with the words “for example” or are set apart from the normative text with class="example", like this:

This is an example of an informative example.

Informative notes begin with the word “Note” and are set apart from the normative text with class="note", like this:

Note, this is an informative note.

9.2. Conformance Classes

Conformance to CSS Image Values and Replaced Content Module Level 4 is defined for three conformance classes:

style sheet
A CSS style sheet.
A UA that interprets the semantics of a style sheet and renders documents that use them.
authoring tool
A UA that writes a style sheet.

A style sheet is conformant to CSS Image Values and Replaced Content Module Level 4 if all of its statements that use syntax defined in this module are valid according to the generic CSS grammar and the individual grammars of each feature defined in this module.

A renderer is conformant to CSS Image Values and Replaced Content Module Level 4 if, in addition to interpreting the style sheet as defined by the appropriate specifications, it supports all the features defined by CSS Image Values and Replaced Content Module Level 4 by parsing them correctly and rendering the document accordingly. However, the inability of a UA to correctly render a document due to limitations of the device does not make the UA non-conformant. (For example, a UA is not required to render color on a monochrome monitor.)

An authoring tool is conformant to CSS Image Values and Replaced Content Module Level 4 if it writes style sheets that are syntactically correct according to the generic CSS grammar and the individual grammars of each feature in this module, and meet all other conformance requirements of style sheets as described in this module.

9.3. Partial Implementations

So that authors can exploit the forward-compatible parsing rules to assign fallback values, CSS renderers must treat as invalid (and ignore as appropriate) any at-rules, properties, property values, keywords, and other syntactic constructs for which they have no usable level of support. In particular, user agents must not selectively ignore unsupported component values and honor supported values in a single multi-value property declaration: if any value is considered invalid (as unsupported values must be), CSS requires that the entire declaration be ignored.

9.4. Experimental Implementations

To avoid clashes with future CSS features, the CSS2.1 specification reserves a prefixed syntax for proprietary and experimental extensions to CSS.

Prior to a specification reaching the Candidate Recommendation stage in the W3C process, all implementations of a CSS feature are considered experimental. The CSS Working Group recommends that implementations use a vendor-prefixed syntax for such features, including those in W3C Working Drafts. This avoids incompatibilities with future changes in the draft.

9.5. Non-Experimental Implementations

Once a specification reaches the Candidate Recommendation stage, non-experimental implementations are possible, and implementors should release an unprefixed implementation of any CR-level feature they can demonstrate to be correctly implemented according to spec.

To establish and maintain the interoperability of CSS across implementations, the CSS Working Group requests that non-experimental CSS renderers submit an implementation report (and, if necessary, the testcases used for that implementation report) to the W3C before releasing an unprefixed implementation of any CSS features. Testcases submitted to W3C are subject to review and correction by the CSS Working Group.

Further information on submitting testcases and implementation reports can be found from on the CSS Working Group's website at http://www.w3.org/Style/CSS/Test/. Questions should be directed to the public-css-testsuite@w3.org mailing list.

9.6. CR Exit Criteria

For this specification to be advanced to Proposed Recommendation, there must be at least two independent, interoperable implementations of each feature. Each feature may be implemented by a different set of products, there is no requirement that all features be implemented by a single product. For the purposes of this criterion, we define the following terms:

each implementation must be developed by a different party and cannot share, reuse, or derive from code used by another qualifying implementation. Sections of code that have no bearing on the implementation of this specification are exempt from this requirement.
passing the respective test case(s) in the official CSS test suite, or, if the implementation is not a Web browser, an equivalent test. Every relevant test in the test suite should have an equivalent test created if such a user agent (UA) is to be used to claim interoperability. In addition if such a UA is to be used to claim interoperability, then there must one or more additional UAs which can also pass those equivalent tests in the same way for the purpose of interoperability. The equivalent tests must be made publicly available for the purposes of peer review.
a user agent which:
  1. implements the specification.
  2. is available to the general public. The implementation may be a shipping product or other publicly available version (i.e., beta version, preview release, or “nightly build”). Non-shipping product releases must have implemented the feature(s) for a period of at least one month in order to demonstrate stability.
  3. is not experimental (i.e., a version specifically designed to pass the test suite and is not intended for normal usage going forward).

The specification will remain Candidate Recommendation for at least six months.


Thanks to the Webkit team, Brad Kemper, Brian Manthos, and Alan Gresley for their contributions to the definition of gradients; to Melinda Grant for her work on ‘object-fit’, ‘object-position’, and ‘image-orientation’; to Michael Day, Håkon Lie, and Shinyu Murakami for ‘image-resolution’; and to L. David Baron, Kang-Hao Lu, Leif Arne Storset, Erik Dahlstrom, and Øyvind Stenhaug for their careful review, comments, and corrections.


Normative references

Bert Bos; et al. Cascading Style Sheets Level 2 Revision 1 (CSS 2.1) Specification. 7 June 2011. W3C Recommendation. URL: http://www.w3.org/TR/2011/REC-CSS2-20110607
Simon Fraser; et al. CSS Transforms. 3 April 2012. W3C Working Draft. (Work in progress.) URL: http://www.w3.org/TR/2012/WD-css3-transforms-20120403/
Håkon Wium Lie; Tab Atkins; Elika J. Etemad. CSS Values and Units Module Level 3. 28 August 2012. W3C Candidate Recommendation. (Work in progress.) URL: http://www.w3.org/TR/2012/CR-css3-values-20120828/
Raphaël Troncy; et al. Media Fragments URI 1.0 (basic). 15 March 2012. W3C Proposed Recommendation. (Work in progress.) URL: http://www.w3.org/TR/2012/PR-media-frags-20120315/
S. Bradner. Key words for use in RFCs to Indicate Requirement Levels. Internet RFC 2119. URL: http://www.ietf.org/rfc/rfc2119.txt
Tantek Çelik; et al. Selectors Level 3. 29 September 2011. W3C Recommendation. URL: http://www.w3.org/TR/2011/REC-css3-selectors-20110929/

Other references

Simon Fraser; et al. CSS 2D Transforms. 15 December 2011. W3C Working Draft. (Work in progress.) URL: http://www.w3.org/TR/2011/WD-css3-2d-transforms-20111215/
Bert Bos; Elika J. Etemad; Brad Kemper. CSS Backgrounds and Borders Module Level 3. 24 July 2012. W3C Candidate Recommendation. (Work in progress.) URL: http://www.w3.org/TR/2012/CR-css3-background-20120724/
Tantek Çelik; Chris Lilley; L. David Baron. CSS Color Module Level 3. 7 June 2011. W3C Recommendation. URL: http://www.w3.org/TR/2011/REC-css3-color-20110607
Håkon Wium Lie; Melinda Grant. CSS3 Module: Paged Media. 10 October 2006. W3C Working Draft. (Work in progress.) URL: http://www.w3.org/TR/2006/WD-css3-page-20061010
Anne van Kesteren. CSSOM. 12 July 2011. W3C Working Draft. (Work in progress.) URL: http://www.w3.org/TR/2011/WD-cssom-20110712/
Thomas Porter; Tom Duff. “Compositing digital images” in: ACM SIGGRAPH Computer Graphics. Volume 18. Issue 3. pp. 253–259. ACM. July 1984. New York, NY, USA. ISSN 0097-8930
Philipp Hoschka. Synchronized Multimedia Integration Language (SMIL) 1.0 Specification. 15 June 1998. W3C Recommendation. URL: http://www.w3.org/TR/1998/REC-smil-19980615
Erik Dahlström; et al. Scalable Vector Graphics (SVG) 1.1 (Second Edition). 16 August 2011. W3C Recommendation. URL: http://www.w3.org/TR/2011/REC-SVG11-20110816/


Property index

Property Values Initial Applies to Inh. Percentages Media
image-orientation <angle> | from-image 0deg all elements yes visual
image-rendering auto | crisp-edges | pixelated auto all elements yes visual
image-resolution [from-image || <resolution>] && snap? 1dppx all elements yes visual
object-fit fill | contain | cover | none | scale-down fill replaced elements no N/A visual
object-position <position> 50% 50% replaced elements no refer to width and height of box itself visual