Clever PNG Optimization Techniques

• By Sergey Chikuyonok
• July 15th, 2009
• Graphics, Optimization, PNG
• 172 Comments

As a web designer you might be already familiar with the PNG image format which offers a full-featured transparency. It’s a lossless, robust, very good replacement of the elder GIF image format. As a Photoshop (or any other image editor) user you might think that there is not that many options for PNG optimization, especially for truecolor PNG’s (PNG-24 in Photoshop), which doesn’t have any. Some of you may even think that this format is “unoptimizable”. Well, in this post we’ll try to debunk this myth.

This post describes some techniques that may help you optimize your PNG-images. These techniques are derived from laborious hours spent on studying how exactly the PNG encoder saves data. We’ll start with some essentials about the PNG format and will then move to advanced optimization techniques.

You may want to take a look at the following related articles:

• Clever JPEG Optimization Techniques

The boring part

Before we dive into image optimization techniques, we have to learn some technical details about the PNG format. Each graphic format has its own advantages and weaknesses; knowing them will allow you to modify original image for better visual quality and compression. This is a key concept behind professional image optimization.

PNG was developed as an open-source replacement of the proprietary GIF format. They have some common features (like indexed color palette), but PNG is much better than GIF in every aspect. It introduced some cool features for image packing and compression, but for us – web-designers and developers – the most important one is the scanline filtering (also known as ‘delta filters’).

Scanline filtering

Here is how it works. For example, we have a 5×5 pixels image with horizontal gradient. Here is a schematic view of this image (each number represents a unique color):

As you can see, all identical colors spread vertically, not horizontally. Such images will have a very poor compression ratio in GIF, because it compresses colors that spread horizontally. Let’s see how this image data can be packed by scanline filtering:

Number 2 before each line represents applied filter, which is “Up” in this case. The “Up” filter sends the message to the PNG decoder: “For the current pixel take the value of the above pixel and add it to the current value.” We have 0 value for lines 2—5 because all pixels in each vertical line have the same color. And such data would be compressed better if the image was relatively large. For example, 15 pixels of value 0 can be written as 0(15) and this is much shorter than fifteen 0′s—this is how compression works in common.

I wrote “can be packed” because in this ideal test case the “Sub” filter (number 1 before each line) will give much better result:

The filter “Sub” sends the message to the decoder: “Take the value of the left pixel and add it to the current value.” In this case, it’s 1. As you may already have guessed, such data will be compressed very effectively.

Scanline filtering is important for us because we can use them: in particular, we can do some image manipulation to make filtering better. There are five filters: None (no filtering), Sub (subtract the left pixel value from the current value), Up (subtract the above pixel value), Average (subtract the average of the left and the upper pixels) and Paeth (substitute the upper, left or upper left pixel value, named after Alan Paeth).

And here’s how these filters affect the image size in comparison with the good ol’ GIF:

GIF, 2568 bytes

PNG, 372 bytes

As you can see, the GIF image is 7 times larger than the same PNG-image.

Image type

Another important thing to know about PNG is image type, the meta-data stored inside the file. As a Photoshop user, you are familiar with PNG-8 (indexed image) and PNG-24 (truecolor image). As a Fireworks user, you may know PNG-32 (truecolor with transparency), which is quite confusing, because Photoshop’s PNG-24 may also store truecolor with transparency. Well, it’s worth knowing that these names are not official, and you won’t find them in PNG specs. For your convenience we’ll use Photoshop’s naming convention of PNG image types in this article.

There are 5 available image types in PNG: Grayscale, Truecolor, Indexed-color, Grayscale with alpha and Truecolor with alpha. There are also two subtypes of indexed-color type (non-official, too): bit transparency (each pixel can be fully transparent or fully opaque) and palette transparency (each pixel can be semi-transparent). In second case each color is stored in palette with its alpha value. Thus, opaque red and 50%-transparent red are two different colors and they take 2 cells inside palette.

The worst thing is that Photoshop can save PNG with only 3 of these types: Indexed-color with bit transparency, Truecolor and Truecolor with transparency. That’s why you may find a lot of opinions that Adobe Fireworks is the best tool for PNG optimization. Personally, I don’t agree with them: Fireworks doesn’t have enough tools for image manipulation, it’s only have slightly more options for saving PNG image, but it’s a topic for another discussion.

This is where utilities such as OptiPNG or pngcrush come in handy. Essentially, these tools do the following:

1. Pick up the best image type for an image (for example, truecolor can be converted to indexed-color if there aren’t too many colors in the image).
2. Pick up best delta filters.
3. Pick up the best compression strategy and, optionally, reduce the color depth.

All these operations do not affect image quality at all, but do reduce the file size of the PNG-images, so I highly recommend you to use such tools every time you save a PNG image.

Now enough with the boring part, let’s do some magic!

1. Posterization

This is a well-known method of the truecolor image optimization. Open up the example image in Photoshop, press the icon in the Layers palette and choose Posterize:

Pick the smallest possible amount of Levels (usually 40 is enough) and save the image:

Original, 84 KB

Posterized, 53 KB

Here is how it works: the posterization simply reduces the amount of colors, converting similar colors to the single one, thus creating posterized regions. This helps to perform a better scanline filtering and achieve a better compression. The downside of this method is color alternation, which is especially visible if you are trying to stitch image with a HTML background:

Original image

Posterized image

2. Dirty Transparency

Take a look at the following images:

75 KB

30 KB

Both of them were saved in Photoshop without any optimization. Even if you do a per-pixel comparison of these images, you won’t notice any difference. But why is the first image 2.5x larger than the second one?

You need a special plugin for Photoshop to see hidden details. It’s called Remove Transparency and available for free download on the PhotoFreebies plugin suite. You have to install it first before proceeding with the next step.

Open both images form the example above in Photoshop and choose Filer ? Photo Wiz ? Remove Transparency. Now you can see the actual pixel data that was saved in the image:

What’s happening? How is it possible to reveal the data from the original image from a single-layered PNG image? Well, it’s quite simple. Each pixel in the truecolor image with alpha is described by four bytes: RGBA. The last one is Alpha, which controls pixel transparency: the value of 0 means fully transparent pixel and 255 means fully opaque. And this means that every pixel (with any RGB value) can be hidden with just Alpha byte set to 0. But this RGB data still exists and, moreover, it prevents PNG encoder from effectively packing and encoding the data stream. Thus, we have to remove this hidden data (fill it with solid black, for example) before saving the image. Here is a quick method how to do this:

1. Open the first image from the example above in Photoshop.
2. Ctrl+click (or ?+click on Mac) on image thumbnail in Layers palette to create a selection, then invert it: Select ? Inverse.

   

3. Switch to Quick Mask mode by pressing Q key:

   

4. We have created a mask for a semi-transparent image, but we need to leave fully transparent pixels only. Choose Image ? Adjustments ? Threshold and move Threshold Level slider to the right, thus leaving fully transparent pixels of the selection:

   

5. Leave Quick Mask mode (press Q key again) and fill the selection with black:

   

6. Invert the seleciton again (Select ? Inverse) and click on the icon in the Layers palette to add mask.

That’s it, now you can save this image in PNG-24 and ensure that the 75 KB image is now 30 KB. By the way, all these steps can be easily recorded into Photoshop’s Action (download the Dirty Transparency Photoshop Action) and reused later with a single keystroke.

You might think about “dirty transparency” as some kind of a bug in image editors: if those image regions can’t be seen and take so much space, why can’t they be removed automatically before saving? Well, this “bug” can be easily turned into a “feature”. Take a look at the following pictures:

5 537 bytes

6 449 bytes

If you remove transparency from those images, you’ll see the following:

Despite the fact that the first image contains more complex image data, it’s 1Kb lighter than the second one, which was optimized as described above. The explanation of this “abnormal” behavior is simple: image data stream in the first example was effectively packed by delta filters, which works better for smooth color transitions (like gradients).

Tech geeks may look at OptiPNG’s output log and ensure that filters were not applied at all for the second image. That’s why I highly recommend you to read The boring part of this article first before using these techniques: if you don’t understand what you’re doing, you can make your image even larger.

The ultimate solution to preserve original image data is to create a mask on the image layer in Photoshop (we’ll come back to this later):

As you can see, Dirty transparency is a very powerful yet very delicate technique. You have to know how and why it works before using it. If you are saving PNG-24 images with transparent areas, the first thing you have to do is to check image data in these areas and make the right decision on clearing or leaving them as is.

3. Split by transparency

Sometimes you have to save image in the “heavy” PNG-24 because of few semi-transparent pixels. You can save extra Kbs if you split such images in two parts — one with solid pixels, the second one with semi-transparent — and save them in appropriate graphic formats. For example, you can save semi-transparent pixels in PNG-24, and solid pixels in PNG-8 or even JPEG. Here is a quick (and recordable for Actions) solution to do this. For our experiments we’ll use this elder Russian iPod ancestor:

PNG-24, 62 KB

1. Ctrl+click/?+click on image thumbnail in Layers palette to create a selection:
2. Go to Channels palette and create new channel from selection:
3. Remove selection (Ctrl+D or ?+D), select the newly created channel and run Threshold (Image ? Adjustments ? Threshold). Move the slider to the very right:
4. We’ve made a mask for selecting dead solid pixels. Now we have to split original layer by this mask. Ctrl+click/?+click on Alpha 1 channel, go to Layers palette, select the original layer and run Layer ? New ? Layer via Cut. As a result, there are two layers with separated solid and semi-transparent pixels.

Now you need to save those two images in separate files: solid pixels in PNG-8, semi-transparent ones in PNG-24. You can apply Posterization technique on semi-transparent pixels layer to make image file even smaller.

PNG-8
128 colors + dithering
17 KB

PNG-24
posterization 35
6 KB

And here is the result for comparison:

Before
63 KB

After
23 KB

This method has an obvious drawback: you get two images instead of one, which may be not so convenient to use (for instance, when making a product catalog in the CMS).

4. Influence masks

Actually, is is not a PNG-specific optimization technique, but demonstration of rarely-used Save for Web properties: Color reduction influence mask and Dithering influence mask.

Sadly, these properties were removed in Photoshop CS4, so you can try this optimization approach only in the pre-CS4 versions (I’m using CS3).

To understand how influence masks works, let’s open this demo image in PS and save it in PNG-8 with the following settings: Color reduction: Adaptive, Dither: No dithering, Colors: 256.

42 KB

The first thing I’ve noticed about this image is a very fuzzy pendulum. It is a very bright spot on the image and it attracts way too much attention. Let’s try to smooth pendulum’s color transitions by setting dithering to 100%:

46 KB

The pendulum looks better now, but we got another problems: image size increased by 4 KB and solid-color background became very noisy:

We can try to get rid of this noise by lowering dithering value, but the image quality may also be reduced.

Based on these problems, let’s try to do the incredible: increase image quality by lowering the number of colors and image size. Influence masks will help us.

Let’s start with the color. Go to Channels palette, create a new channel and name it color. We’ve already determined that the pendulum is our top priority region to improve image quality, so we need to draw a white circle right on its place (you can enable RGB channel for better precision).

Go to Save for Web dialog and set the following properties: Color reduction: Adaptive, Dither: No, Colors: 128 (as you can see, we reduced number of colors from 256 to 128). Now we have to select an influence mask: click on the near Color reduction list and select the color channel from drop-down list: Now our image looks as follows:

You can see the influence mask in action: the pendulum looks perfect, but the other parts of image look really bad. By setting influence mask, we said to Photoshop: “Look, mate, the pendulum is very important part of image so try to preserve as much colors in this area as possible”. Influence mask works exactly the same as regular transparency mask: white color means highest priority in corresponding image region, black color means lowest priority. All intermediate shades of gray affect on image proportionally.

The pendulum now takes the highest color priority, so we have to lower the intensity of white circle to leave more colors for other areas. Close Save for Web dialog, go to Channels palette, select color channel and open Levels dialog (Image ? Adjustments ? Levels). Set the maximum output level to 50 to lower the white color intensity:

Try to save for Web again with the same properties:

Looks better now, but now we’ve got problems in other image areas:

I think you already understand how influence masks works: you provide Photoshop with some clues about important image areas with different shades of gray. With trials and errors I’ve got the following color mask (you can copy it and apply to the image):

Dithering influence mask works exactly the same, but instead of colors, it affects the dithering amount of different image areas. Lighter color means more dithering. This is a very useful feature, because dithering creates irregular pixel patterns which hinders the PNG compressor to use delta filters. You can determine the exact areas where dithering must be applied while leaving other areas intact, thus gaining better compression of image data.

My dithering channel looks like this:

Applying both color and dithering influence channels with the same optimization settings (Adaptive, 128 colors):

Pretty good for 128 colors, isn’t it? Let’s do some finishing touches: set colors to 180 and max dithering to 80%. And here is our final result compared to original, non-optimized version:

256 colors, no dithering, non-optimized
42 KB

180 colors, optimized
34 KB

Please stay tuned (RSS, Twitter) for the second part of the article where we’ll cover further techniques; we’ll talk about grayscale images, using less colors, lowering details and discuss further tips for PNG usage and optimization as well as the PNG optimization in practice.

Related posts

You may want to take a look at the following related articles:

• Clever JPEG Optimization Techniques