Introduction¶

Welcome to OpenImageIO!

I kinda like “Oy-e-oh” with a bit of a groaning Yiddish accent, as in
“OIIO, did you really write yet another file I/O library?”

Dan Wexler

Overview¶

OpenImageIO provides simple but powerful ImageInput and ImageOutput APIs that abstract the reading and writing of 2D image file formats. They don’t support every possible way of encoding images in memory, but for a reasonable and common set of desired functionality, they provide an exceptionally easy way for an application using the APIs support a wide — and extensible — selection of image formats without knowing the details of any of these formats.

Concrete instances of these APIs, each of which implements the ability to read and/or write a different image file format, are stored as plugins (i.e., dynamic libraries, DLL’s, or DSO’s) that are loaded at runtime. The OpenImageIO distribution contains such plugins for several popular formats. Any user may create conforming plugins that implement reading and writing capabilities for other image formats, and any application that uses OpenImageIO would be able to use those plugins.

The library also implements the helper class ImageBuf, which is a handy way to store and manipulate images in memory. ImageBuf itself uses ImageInput and ImageOutput for its file I/O, and therefore is also agnostic as to image file formats. A variety of functions in the ImageBufAlgo namespace are available to perform common image processing operations on ImageBuf’s.

The ImageCache class transparently manages a cache so that it can access truly vast amounts of image data (thousands of image files totaling hundreds of GB to several TBs) very efficiently using only a tiny amount (tens of megabytes to a few GB at most) of runtime memory. Additionally, a TextureSystem class provides filtered MIP-map texture lookups, atop the nice caching behavior of ImageCache.

Finally, the OpenImageIO distribution contains several utility programs that operate on images, each of which is built atop this functionality, and therefore may read or write any image file type for which an appropriate plugin is found at runtime. Paramount among these utilities oiiotool, a command-line image processing engine, and iv, an image viewing application. Additionally, there are programs for converting images among different formats, comparing image data between two images, and examining image metadata.

All of this is released as “open source” software using the very permissive “BSD 3-clause” license. So you should feel free to use any or all of OpenImageIO in your own software, whether it is private or public, open source or proprietary, free or commercial. You may also modify it on your own. You are encouraged to contribute to the continued development of OpenImageIO and to share any improvements that you make on your own, though you are by no means required to do so.

Simplifying Assumptions¶

OpenImageIO is not the only image library in the world. Certainly there are many fine libraries that implement a single image format (including the excellent libtiff, libjpeg, and OpenEXR that OpenImageIO itself relies on). Many libraries attempt to present a uniform API for reading and writing multiple image file formats. Most of these support a fixed set of image formats, though a few of these also attempt to provide an extensible set by using the plugin approach.

But in our experience, these libraries are all flawed in one or more ways: (1) They either support only a few formats, or many formats but with the majority of them somehow incomplete or incorrect. (2) Their APIs are not sufficiently expressive as to handle all the image features we need (such as tiled images, which is critical for our texture library). (3) Their APIs are too complete, trying to handle every possible permutation of image format features, and as a result are horribly complicated.

The third sin is the most severe, and is almost always the main problem at the end of the day. Even among the many open source image libraries that rely on extensible plugins, we have not found one that is both sufficiently flexible and has APIs anywhere near as simple to understand and use as those of OpenImageIO.

Good design is usually a matter of deciding what not to do, and OpenImageIO is no exception. We achieve power and elegance only by making simplifying assumptions. Among them:

• OpenImageIO only deals with ordinary 2D images, and to a limited extent 3D volumes, and image files that contain multiple (but finite) independent images within them. OpenImageIO’s support of “movie” files is limited to viewing them as a sequence of separate frames within the file, but not as movies per se (for example, no support for dealing with audio or synchronization).

• Pixel data are presented as 8- 16- or 32-bit int (signed or unsigned), 16- 32- or 64-bit float. NOTHING ELSE. No $$<8$$ bit images, or pixel value boundaries that aren’t byte boundaries. Files with $$<8$$ will appear to the client application as 8-bit unsigned grayscale images.

• Only fully elaborated, non-compressed data are accepted and returned by the API. Compression or special encodings are handled entirely within an OpenImageIO plugin.

• Color space is by default converted to grayscale or RGB. Non-spectral color models, such as XYZ, CMYK, or YUV, are converted to RGB upon reading. (There is a way to override this and ask for raw pixel values.)

• All color channels can be treated (by apps or readers/writers) as having the same data format (though there is a way to deal with per-channel formats for apps and readers/writers that truly need it).

• All image channels in a subimage are sampled at the same resolution. For file formats that allow some channels to be subsampled, they will be automatically up-sampled to the highest resolution channel in the subimage.

• Color information is always in the order R, G, B, and the alpha channel, if any, always follows RGB, and Z channel (if any) always follows alpha. So if a file actually stores ABGR, the plugin is expected to rearrange it as RGBA.

It’s important to remember that these restrictions apply to data passed through the APIs, not to the files themselves. It’s perfectly fine to have an OpenImageIO plugin that supports YUV data, or 4 bits per channel, or any other exotic feature. You could even write a movie-reading ImageInput (despite OpenImageIO’s claims of not supporting movies) and make it look to the client like it’s just a series of images within the file. It’s just that all the nonconforming details are handled entirely within the OpenImageIO plugin and are not exposed through the main OpenImageIO APIs.

Historical Origins¶

OpenImageIO is the evolution of concepts and tools I’ve been working on for two decades.

In the 1980’s, every program I wrote that output images would have a simple, custom format and viewer. I soon graduated to using a standard image file format (TIFF) with my own library implementation. Then I switched to Sam Leffler’s stable and complete libtiff.

In the mid-to-late-1990’s, I worked at Pixar as one of the main implementors of PhotoRealistic RenderMan, which had display drivers that consisted of an API for opening files and outputting pixels, and a set of DSO/DLL plugins that each implement image output for each of a dozen or so different file format. The plugins all responded to the same API, so the renderer itself did not need to know how to the details of the image file formats, and users could (in theory, but rarely in practice) extend the set of output image formats the renderer could use by writing their own plugins.

This was the seed of a good idea, but PRMan’s display driver plugin API was abstruse and hard to use. So when I started Exluna in 2000, Matt Pharr, Craig Kolb, and I designed a new API for image output for our own renderer, Entropy. This API, called “ExDisplay,” was C++, and much simpler, clearer, and easier to use than PRMan’s display drivers.

NVIDIA’s Gelato (circa 2002), whose early work was done by myself, Dan Wexler, Jonathan Rice, and Eric Enderton, had an API called “ImageIO.” ImageIO was much more powerful and descriptive than ExDisplay, and had an API for reading as well as writing images. Gelato was not only “format agnostic” for its image output, but also for its image input (textures, image viewer, and other image utilities). We released the API specification and headers (though not the library implementation) using the BSD open source license, firmly repudiating any notion that the API should be specific to NVIDIA or Gelato.

For Gelato 3.0 (circa 2007), we refined ImageIO again (by this time, Philip Nemec was also a major influence, in addition to Dan, Eric, and myself 1). This revision was not a major overhaul but more of a fine tuning. Our ideas were clearly approaching stability. But, alas, the Gelato project was canceled before Gelato 3.0 was released, and despite our prodding, NVIDIA executives would not open source the full ImageIO code and related tools.

1

Gelato as a whole had many other contributors; those I’ve named here are the ones I recall contributing to the design or implementation of the ImageIO APIs.

After I left NVIDIA, I was determined to recreate this work once again – and ONLY once more – and release it as open source from the start. Thus, OpenImageIO was born. I started with the existing Gelato ImageIO specification and headers (which were BSD licensed all along), and made further refinements since I had to rewrite the entire implementation from scratch anyway. I think the additional changes are all improvements.

Over the years and with the help of dozens of open source contributors, OpenImageIO has expanded beyond the original simple image format input/output to encompass a wide range of image-related functionality. It has grown into a foundational technology in many products and tools, particularly for the production of animation and visual effects for motion pictures (but also many other uses and fields). This is the software you have in your hands today.

Acknowledgments¶

OpenImageIO incorporates, depends upon, or dynamically links against several other open source packages, detailed below. These other packages are all distributed under licenses that allow them to be used by OpenImageIO. Where not specifically noted, they are all using the same BSD license that OpenImageIO uses. Any omissions or inaccuracies in this list are inadvertent and will be fixed if pointed out. The full original licenses can be found in the relevant parts of the source code.

OpenImageIO incorporates, distributes, or contains derived works of:

OpenImageIO Has the following build-time dependencies (using system installs, referencing as git submodules, or downloading as part of the build), including link-time dependencies against dynamic libraries: