High dynamic range (HDR) images cannot be displayed directly on conventional output devices like consumer monitors or printers since these devices cover a much smaller dynamic range than the HDR image. Therefore, HDR images have to be converted into low dynamic range (LDR) images before publishing them on a website or creating prints.

Tone mapping algorithms can be used to automate the process of converting an HDR image to an LDR image by compressing the dynamic range of the HDR image. Various tone mapping algorithms have been proposed in the scientific literature, and several tools exist that implement these algorithms. The goal of this website is to compare existing implementations and readily available tools from a practitioner's rather than a scientest's point of view. In contrast to the objective scientific comparisons presented in the scientific literature, the comparison presented here is highly subjective based on my personal experience. I perfectly know that possibly not everybody will agree with my interpretations. However, since I will also present the images resulting from tone mapping, everyone can make up his own mind.

Scenario

Tone mapping obviously only makes sense if the contrast of the original HDR image is high enough to prevent it from being mapped directly to an LDR image. In order to test the ability of the algorithms to compress an HDR image's dynamic range efficiently, I selected an HDR image that I consider to be difficult in terms of tone mapping since it contains very bright and dark parts. This Figure depicts the test scenario. Since the original HDR image cannot be shown directly on common displays, this Figure shows the scene as a single (LDR) image as it was taken with a common SLR camera (Olympus E-300). This LDR image serves as reference since it represents an image that would be the result of common photography not using HDR and tone mapping. The HDR image (38 MByte) of this scene has been created from five LDR images taken with the same camera at different exposures (aperture 8; shutter speeds 1/4, 1/15, 1/60, 1/250, 1/1000). The tool FDRTools Basic Version 2.2 has been used to create the HDR image from these images.

Single (LDR) image taken with camera (click on image to enlarge)

The chosen scenario contains two critical parts. The white church in the bright sunlight, and the dark steel framework of the bridge in the shadow. The Figure shows quite nicely that an ordinary LDR camera image does not show any details of the dark side of the bridge. If we adjusted the exposure to the dark parts, the church and sky would be heavily over-exposed. In contrast to this LDR image, the HDR image contains information of both dark and bright parts of the image. Next, let us see whether the tone mapping algorithms can produce an LDR image from this HDR image that shows both parts with all details.

Evaluation Criteria

Before the comparison is presented, some criteria for comparing the output of the tone mapping algorithms are required. As said, this comparison does not strive to be strictly objective. Still I want to explain what I personally consider to be important from my subjective point of view.

Dynamic Range Compression

The primary goal of tone mapping algorithms is to compress the dynamic range of the original HDR image to fit the much smaller dynamic range covered by an LDR image while still preserving the details of the original image. In plain words, this means that bright parts of the image should not be burned out to white areas and dark parts should not drown in black areas. Details should be visible in both dark and bright parts of the image. That is, the algorithm should support high dynamic range compression.

Natural Look

Tone mapping manipulates the original HDR image. This might result in an image losing its natural look that would be perceived in reality by a human observer. Although some people use tone mapping for creating artistic and surreal images, my personal goal is an image that comes as close as possible to the real situation.

To some degree, natural look and high dynamic range compression are contradicting goals. It is quite hard to define objectively what a natural look really means, and I refer all readers interested in an objective comparison to the scientific literature. In my comparison, the natural look of the image is judged by me by simply having a close look at the image. However, to make it at least a little bit more concrete, here are some things that should be avoided in my opinion:

High saturation, i.e., the image should not look like an image taken from a comic.
Overly sharp images.
Halos, i.e., dark areas around bright areas or vice versa.

Evaluation

The following images are the result of different tone mapping algorithms applied to the HDR image. In order to get images that can be compared visually, the black and white point as well as the gamma value of the image have been adjusted manually. No other post processing has been applied to the images since the goal of this comparison is to evaluate the implementations of the tone mapping algorithms rather than the one who post processes the images.

All tools use a set of parameters to adjust the implemented algorithms to the particular image. In order to create the images presented here, those parameters have been used that led to the best results, i.e., the best trade-off of natural look and dynamic range compression.

Drago Algorithm (pfstools)

The open source HDR toolkit pfstools provides implementations of different tone mapping algorithms. Pfstools is a command line application, although graphical frontends exist (qpfstmo, qtpfsgui). Pfstools is free and open source.

This Figure shows the result of the algorithm of Drago et al. Although this algorithm is quite fast and can be configured easily with only one parameter, it is not able to achieve the necessary high dynamic range compression in order to show details in both bright and dark parts of the test image. Moreover, the image suffers from a noticable yellowish color cast and oversaturated colors.

Drago Algorithm (pfstools) (click on image to enlarge)

Durand Algorithm (pfstools)

The algorithm of Durand et al. (pfstools implementation) produces a naturally looking image with correct colors and no oversaturation (see this Figure), although its runtime is significantly higher, and parameters are more complex to adjust than for the Drago algorithm. Moreover, the image does not show artifacts such as halos around dark or bright area. The dynamic range compression is good, although it could be a little bit higher in my opinion. The image shows details of the dark bridge without burning out bright details. Overall, a good result.

Durand Algorithm (pfstools) (click on image to enlarge)

Durand Algorithm (Implementation by Durand)

Besides the pfstools implementation, there exists another free open source implementation of Durand's algorithm by Durand himself. This implementation is faster and much easier to configure. It requires only one parameter that in many cases can be set to the default value to achieve good results. This alleviates the fact that this implementation is only provided as command line tool without GUI.

In terms of dynamic range compression, Durand's implementation seems to have a slighly better performance. The implementation by Durand produces an image with slightly higher saturation, however, the image is not oversaturated. Overall, a solid and convincing performance.

Durand Algorithm (Implementation by Durand) (click on image to enlarge)

Mantiuk Algorithm (pfstools)

Pfstools provides implementations of two tone mapping algorithms by Mantiuk et al., called contrast mapping and contrast equalization. Since contrast equalization did not terminate correctly, only the results of the contrast mapping algorithm are presented in this Figure.

On the one hand, Mantiuk's algorithm is able to bring out details of the dark bridge and shows good dynamic range compression without the dreaded halo artifacts. Moreover, it does not lead to oversaturated colors. On the contrary, colors are slightly undersaturated, however, this can be correctly easily during post processing.

On the other hand, the algorithm tends to produce overly sharp images. Although there is a parameter to control this behaviour, it is a trade-off of dynamic range compression and sharpness.

Mantiuk Algorithm (pfstools) (click on image to enlarge)

Fattal Algorithm (pfstools)

This Figure shows the result of Fattal's tone mapping algorithm as implemented by pfstools. This algorithms exhibits very high dynamic range compression. No other algorithm in this comparison shows the details of the bridge so clearly.

However, this high compression comes at the cost of significant artifacts such as strong halos, and color casts. For instance, if you have a closer look at the house behind the steel framework of the bridge on the left side, you will notice white areas where the house is supposed to be. Moreover, around the bright white church there is a black halo. Overall, this can hardly be called a realistic image.

Actually, the pfstools developers try very hard to alleviate these drawbacks by adding more parameters to control the algorithm. However, currently no silver bullet has been found, and the algorithm becomes even more complex for the user because of its many parameters.

Fattal Algorithm (pfstools) (click on image to enlarge)

Receptor Algorithm (FDRTools Basic)

In contrast to the open source project pfstools, FDRTools is a closed source product. Besides the commercial version FDRTools Advanced, a freeware version FDRTools Basic is available that was used to create the results presented in this section. FDRTools Basic implements two tone mapping algorithms. A basic algorithm and the more sophisticated receptor algorithm. The image shown in this Figure is the result of the receptor algorithm.

Although colors are correct and not oversaturated, the achieved dynamic range compression is low, comparable to Drago's algorithm. So, although I think FDRTools Basic is a great tool for creating HDR images from multiple LDR images – the HDR image used for this comparison has been created with FDR Tools –, other tools provide more effective tone mapping algorithms, at least compared to the freeware version FDRTools Basic.

Receptor Algorithm (FDRTools Basic) (click on image to enlarge)

Compressor Algorithm (FDRTools Advanced)

The commercial version of FDRTools (FDRTools Advanced) comes with a more powerful tone mapping operator called compressor algorithm. The image created with the compressor algorithm shows much higher dynamic range compression than the receptor algorithm of FDRTools Basic. The details of the dark bridge and the white church a clearly visible whithout the dreaded halo artifacts. The saturation seems to be higher than for the Durand algorithm although this can be adjusted with a parameter of the compressor algorithm.

Compressor Algorithm (FDRTools Advanced) (click on image to enlarge). Note: This image contains a watermark that is not a result of tone mapping.

Detail Enhancer Algorithm (Photomatix)

The commercial tool Photomatix implements two tone mapping algorithms called tone compressor and detail enhancer. Compared to detail enhancer, the tone compressor algorithm is less powerful. Actually, for the presented HDR image, it was not possible to produce a satisfactory result. Therefore, the remainder of this section is focussed on the detail enhancer algorithm.

This Figure shows the result of the detail enhancer algorithm of Photomatix using conservative parameters. It has a natural look without noticable artifacts. The dynamic range compression is good and reveals details of the dark bridge comparable to Durand's algorithm, however, the compression is a little bit smaller than the compression achieved by the compressor algorithm of FDRTools Advanced.

Detail Enhancer Algorithm (Photomatix); conservative parameters (click on image to enlarge). Note: This image contains a watermark that is not a result of tone mapping.

Even higher dynamic range compression can be achieved using more aggressive parameters as shown in this Figure). Although details of the dark bridge are much clearer now, this comes at the cost of visible artifacts with halos around the bridge and overly sharp details.

Detail Enhancer Algorithm (Photomatix); aggressive parameters (click on image to enlarge). Note: This image contains a watermark that is not a result of tone mapping.

Tone Mapper (EasyHDR)

The commercial tool EasyHDR provides a tone mapping algorithm that according to the tool's website is tailored to realistic images. In fact, the image created with EasyHDR (see this Figure) backs this claim up. This image has realistic colors without being overly sharp or saturated and without halo artifacts. The dynamic range compression is still good, comparable to Durand's algorithm or Photomatix (with conservative parameters) although a little bit smaller than the compression achieved by the compressor algorithm of FDRTools Advanced.

Tone Mapper (EasyHDR) (click on image to enlarge). Note: This image contains a watermark that is not a result of tone mapping.

Conclusion

Looking at the results presented above, we can conclude that high dynamic range compression and a natural look can be achieved at the same time indeed. In particular, the tone mapping algorithms of Durand (pfstools and Durand's implementation), the detail enhancer algorithm (Photomatix) if used with conservative parameters, the compressor algorithm of FDRTools Advanced, and the tone mapping operator of EasyHDR were able to produce convincing results. Photomatix seems to be most flexible since its tone mapping algorithm can be parametrized to produce everything from natural to surreal images. The algorithms of FDRTools Advanced, EasyHDR, and Durand are mainly aiming for realistic images, where they do a slightly better job than Photomatix in my opinion. Overall, the compressor tone mapping algorithm of FDRTools Advanced seems to achieve the best trade-off of high contrast compression and realistic images.

If one likes the sharp look of Mantiuk's algorithm, this is also a good choice in terms of dynamic range compression. For the people preferring surreal images rather than realistic images, I would recommend the detail enhancer algorithm of Photomatix with more aggressive parameter settings.

This comparison also shows that good tone mapping algorithms are available at no cost (pfstools and Durand's implementation are free and open source), although commercial tools such as Photomatix, FDRTools, or EasyHDR are easier to use, mainly because a GUI with real time preview is available with these commercial tools that helps significantly in adjusting the parameters.

However, as pointed out several times in this article: this comparison is highly subjective based on my personal experience and preferences. For a fast visual comparison, you can use the following images to show the final results of the different algorithms side-by-side. Select the algorithm you want to compare for the left and right image in order to display the resulting images.

References

F. Drago, K. Myszkowski, T. Annen, and N. Chiba: Adaptive logarithmic mapping for displaying high contrast scenes. In Computer Graphics Forum 22(3), 2003.
F. Durand and J. Dorsey: Fast Bilateral Filtering for the Display of High-Dynamic-Range Images. In ACM Transactions on Graphics, 2002.
Rafal Mantiuk, Karol Myszkowski, Hans-Peter Seidel: A Perceptual Framework for Contrast Processing of High Dynamic Range Images. In ACM Transactions on Applied Perception 3(3), 2006
R. Fattal, D. Lischinski, and M. Werman: Gradient Domain High Dynamic Range Compression. In ACM Transactions on Graphics, 2002.


Drago (pfstools) Durand (pfstools) Durand (Durand) Mantiuk (pfstools) Fattal (pfstools) Receptor (FDRTools Basic) Compressor (FDRTools Advanced) Detail Enhancer; conservative parameters (Photomatix) Detail Enhancer; aggressive parameters (Photomatix) Tone Mapper (EasyHDR)	Drago (pfstools) Durand (pfstools) Durand (Durand) Mantiuk (pfstools) Fattal (pfstools) Receptor (FDRTools Basic) Compressor (FDRTools Advanced) Detail Enhancer; conservative parameters (Photomatix) Detail Enhancer; aggressive parameters (Photomatix) Tone Mapper (EasyHDR)