Color Quantization For Film


RenderDotC's default behavior is to generate 8 bit linear data.  This means that each channel (red, green, blue, and alpha) spans from nominal black at 0 to nominal white at 255.  While this is often acceptable, RenderDotC is capable of producing higher quality output.  When rendering for film, the 8 bit model may run out of steam.  This document explains why rendering for film is more demanding on color quantization (especially when composited with live action) and how to achieve superior results with RenderDotC.


Background

For a thorough and accurate description of film and the 10 bit digital negative (Cineon) format, see these original documents:
Conversion of 10-bit Log Film Data To 8-bit Linear or Video Data by Kodak Picture & Television Imaging
Grayscale Transformations of Cineon Digital Film Data by Cinesite Digital Film Center
The following is a layman's explanation:

Film has a tremendous dynamic range [4].  It starts at black and must be overexposed mercilessly before it becomes totally saturated and can get no whiter.  Instead of considering the total possible range of film, we start by focusing on the normal range of intensities that might occur when filming an indoor or outdoor scene, from a reference black to a reference white.  Cameras are sometimes calibrated by holding benchmark gray or white cards in front of the lens.  Such cards are of a known intensity expressed as a percentage of 100% reference white.  The 18% gray card is used in film photography and television cameras are calibrated against the 90% white card [2].

The Cineon format uses ~1% black card as reference black and 90% white card as reference white [1].  On film, total saturation is perhaps 20 times brighter than reference white.  This additional range is called "headroom" and can be used to capture specular highlights on water or chrome.  Cineon's digital negative also captures the extended headroom.  A logarithmic scale is employed to focus the numeric precision where it is needed most on the darker shades, sacrificing precision in the headroom.  Cineon uses 10 bits to avoid contouring (mach bands) in digital images.  Therefore, this format is known as "10-bit log".


RenderMan Limitations

The RenderMan Interface Specification 3.1 [5] requires that the default color quantization be 8 bit linear.  RenderDotC honors this requirement.  One may change the color quantization with RiQuantize:
Quantize type one min max ditheramplitude
The arguments to RiQuantize are as follows: One limitation of the RenderMan model of quantization is that only linear mapping is possible.  Creating a 10 bit log scale such as Cineon must be done as a post processing step.

Another limitation is that colors and alpha must have identical quantization parameters [3].  As we shall see later, RenderDotC overcomes this limitation with an implementation specific extension to the RenderMan standard.


Quantization in Practice

When rendering for film, the first problem encountered is that 8 bits are not enough [3].  Artifacts such as mach bands may appear.  The solution is simple enough.  Use 16 bits:
Quantize "rgba" 65535 0 65535 0.5      # 16 bit integer
Or even 32 bit floating point:
Quantize "rgba" 0 0 0 0                # 32 bit floating point
[Note that min, max, and ditheramplitude are meaningless when using floating point output.  Here, they are arbitrarily set to all zeros.]

When quantizing to integers, RenderDotC looks only at max to determine how many bits per channel to use.  Possible values are 1, 2, 4, 8, or 16 bits per channel.  The smallest number of bits that will accommodate max is selected.  It makes sense to choose a value for max that uses all of the bits.  In the example above, we wanted to use 16 bits so we set max = 216 - 1 = 65535.  This is the largest possible unsigned 16 bit number.

The standard (but shortsighted) approach that is often taken when recording computer generated images (CGI) to film is to align 0 with reference black and max with reference white.  This works reasonably well except that the extended headroom of the film goes unused.  Bright highlights in the CGI just aren't as bright as they should be.  If the CGI is mixed with live action, it may look dull and flat by comparison [4].

Here's where the one parameter of RiQuantize comes into play.  We can set one to some value less than max, align one with reference white, and leave the range from one to max for the extended headroom.  If an object in the scene is a fully illuminated white object, the shader will return Ci = color(1.0) and reference white will be met.  A specular highlight off of chrome may produce an even more intense color such as color(5.0).  Be careful that the shader does not arbitrarily clamp all colors to 1.0.  Otherwise, the headroom will never be exercised.

What's a good value for one?  The perfect value for covering the same extended range as 10 bit log is about 4829.  This is close enough to 212 - 1 = 4095 that we may substitute it for the convenience of nice round numbers:

Quantize "rgba" 4095 0 65535 0.5
Don't make the mistake of choosing 1023 just because that sounds like a good value for 10 bit log.  The RenderMan quantization space remains linear.  10 bits in linear space has nothing to do with 10 bit log.  Choosing 1023 results in leaving more headroom than can be captured on film at the cost of reduced precision below reference white.


Texture Files

When using an extended linear range, texture files (especially environment maps) should be stored in the same space.  If the image was created with RenderDotC, then reference white is automatically stored as the value of one and is automatically transferred to the texture/environment maps by texdc (or RiMakeTexture).  The texture() and environment() functions used by the shader then convert the texel to floating point by dividing it by reference white.  For a texel at reference white of 4095, texture() will return 4095 / 4095 = 1.0, exactly the desired result.  A texel at the maximum value of 65535 results in a color of 16.0, well above reference white and into the range reserved for the headroom.

For images not created by RenderDotC, reference white may be explicitly set when building the texture or environment maps.  For more information, see the documentation on the refwhite option.


A RenderDotC Extension

One problem with the procedure described above is that alpha goes through the same quantization procedure as color [3].  It doesn't make sense for alpha to have headroom.  We really just want a simple, 16 bit linear alpha channel.  The RenderMan Specification [5] provided no mechanism for separating RGB from A in RiQuantize.  RenderDotC provides an extension to do just that:
Quantize "rgb" 4095 0 65535 0.5        # 16 bit linear with headroom
Quantize "a" 65535 0 65535 0.5         # Straight 16 bit linear
RenderDotC still recognizes the type "rgba" and simply applies the quantization to both RGB and A.


References

1. Conversion of 10-bit Log Film Data To 8-bit Linear or Video Data, Kodak Picture & Television Imaging, 1995
2. Grayscale Transformations of Cineon Digital Film Data, Cinesite Digital Film Center, 1993
3. Advanced RenderMan: Beyond the Companion, SIGGRAPH 1998 Course Notes, Antoine Durr (unpublished)
4. Personal Communication, Kim Libreri, 1998
5. The RenderMan Interface Specification Version 3.1, Pixar Animation Studios, 1989
 


Copyright © 1999-2006 Dot C Software, Inc. All rights reserved.
Dot C Software, Inc., 182 Kuuhale St., Kailua, HI 96734
(808) 262-6715 (voice) (808) 261-2039 (fax)
The RenderMan® Interface Procedure and RIB Protocol are:
Copyright © 1988, 1989, Pixar.  All rights reserved.
RenderMan® is a registered trademark of Pixar.