High Dynamic Range Images

Please update to the following new link -- this page is outdated:

http://www.anyhere.com/gward/hdrenc/hdr_encodings.html

The pixel values in most digital images are limited to a useful dynamic range of about 100:1, or two orders of magnitude. Although this corresponds well to the capabilities of most computer graphics displays, it falls short of what people are actually capable of perceiving in real environments, and falls far short of the range of real-world luminance values.

The human eye is capable of dynamically adjusting to a portion of a scene that is 1/10,000 as bright as the brightest point in the field of view. Thus, we would like to be able to represent a range of four orders of magnitude on an ideal display device. We are currently very far from this goal, but it is conceivable that future VR displays will achieve something close to this dynamic range. High dynamic range displays are desirable for testing visual performance and comfort in critical, challenging, or novel situations.

Even if we cannot display the full range of perceivable luminances, we may wish to apply some advanced tone-mapping algorithm to represent this larger range on a more limited device. Such algorithms have been described by Larson et al and NASA.

If we lose the original information to an inadequate image format, we lose both display options.

High Dynamic Range Image Formats

There are at least three high dynamic range image formats freely available. We are promoting our own format, which we believe is the most accurate and general of the three.

Pixar's Log Format

For many years, Pixar has recognized the need for dynamic range in film recording, so they developed a 33-bit/pixel log encoding for RGB values, which is implemented as part of Sam Leffler's TIFF library. This code depends on the ZIP (deflate) library. The encoding covers about 3.5 orders of magnitude, with 0.4% relative accuracy.

Radiance's RGBE Format

Since it's inception in 1987, the Radiance rendering system has used a 32-bit/pixel floating point format, which is described in Greg Ward's chapter entitled "Real Pixels," in Graphics Gems IV, edited by James Arvo, Academic Press, 1994. Additional information and code may be found on the Radiance web site. The encoding covers about 76 orders of magnitude with 1% relative accuracy.

SGI's LogLuv Format

Through a recent addition to Sam Leffler's TIFF library, version 3.4beta037 includes a new codec for a 32-bit/pixel logLuv encoding. Compared to the Radiance and Pixar encodings, this has the advantage of covering the full gamut of perceivable colors in imperceptible steps. The luminance range covers 38 orders of magnitude with 0.3% relative accuracy, and chroma errors never reach the the visible threshold. See the following web pages for the format proposal, describing motivations and goals, and details about the implementation, with example code and images.

Questions, comments, and chocolate should be sent to:

Gregory Ward Larson (gregl@sgi.com)
Silicon Graphics, Inc.

Last Update: September 8, 1997.