Please add your experience with the different algorithms into this wiki to help other users to achieve good results in a faster time.
Unless stated otherwise it is assumed that the HDR image is created with the default values, thus only the tone mapping values have to be changed.
If you don't like the result of a specific tone mapping operator, please keep in mind that after the tone mapping step you can still use tools like GIMP to post process the resulting image. For example, you can still fix the brightness, change the gamma or the levels, and so on.
For a detailed description of the operators, and things like "Alpha", "beta", "range kernel sigma" and so on refer back to the respective paper's equations (you can find them at):
http://www.mpi-inf.mpg.de/resources/tmo
Ashikmin
ashikhmin02 eq2 local=0.646465
When to use this algorithm
Pictures can be very high in contrast and very colourful (if all sliders are moved quite far to the left side) or very washed out. The washed out reminds on a mix of water colours and charcoal drawing.
Parameters in detail
Simple (Default not set)
It looks like that using the simple button makes it easier to achieve more realistic pictures. It seems as well that "simple" mode is quite close to "Equation 4" mode if not the simple button is set.
Drago
drago03 bias=0.85
When to use this algorithm
This algorithm is intended to imitate the human eye's response, and is useful when a true tone result is desired. It is a global spatially uniform operator: at the beginning, it calculates the average luminance of the image and, using this value and the external parameter "bias", it creates a non-linear logarithmic function that is applied to each pixel separately, without considering the neighbouring pixels. This way the algorithm is fast and it will not change the overall look of the image, but it is less adaptive and the details will be less visible compared to other algorithms. It can be used on noisy images without any particular problem because details are not amplified more than other parts of the image.
Parameters in detail
See the web page at http://www.mpi-inf.mpg.de/resources/tmo/logmap/ for technical details.
Bias (default 0.85) +++
The Bias expands or contracts the dynamic range of the tone mapped image. It changes the base of the logarithm that is used in the tone-mapping process. If it is smaller, the dynamic range will be compressed more and the overall look of the LDR image will be brighter. After several attempts, the authors have found 0.85 to be the best value to achieve realistic and clear results.
Durand
durand02 spatial=8 range=0.4 base=5
When to use this algorithm
This algorithm produces the most realistic pictures. No extreme effects, but very nice output with lots of details in the picture. This algorithm needs a lot of time to calculate the output.
Parameters in detail
Base Contrast (Default 5)
Found this value usually too high and reduce it between 2 to 3. The lower the value the lower is the contrast, the picture gets darker. This can be important since it seems that light spots are too bright.
Range Kernel Sigma (Default 0.4)
Increasing this value creates a halo around lights, somewhat similar to soft focus filter.
Post tone mapping gamma adjustments (Default 1)
The higher the value here the lower the contrast will be. Seems to have a similiar effect like the base contrast setting.
Fattal
fattal02 alpha=0.1 beta=0.8 saturation=1
When to use this algorithm
This algorithm is the one used by most users to produce pictures with a stunning effects. Usually these pictures are not realistic, but can be very impressive. Note that, if your images contain noise (e.g. you took them using a high ISO value), it will be heavily amplified (they are considered details), becoming much more evident. You'd better not to use this algorithm on noise images.
Parameters in detail
The algorithm works manipulating the gradient field of the luminance image by attenuating the magnitudes of large gradients. Very large changes occuring on the edges are reduced; on the other hand, smaller details are amplified thus becoming more evident. Changing the value of pre-gamma causes only minor changes to the output.
Alpha (Default 0.1)
The parameter alpha is a threshold: details whose derivative is less than alpha are amplified, those whose derivative is greater than alpha are decreased. In other words, decreasing alpha the quantity of details that are made more evident is decreased. Increasing alpha will make the details more evident, but the picture will look more unrealistic and noisy (because noise will be amplified as well). The authors said to have used the value of 0.1 in each test with best results.
Beta (Default 0.8)
The parameter beta expresses how much the algorithm will be effective. Setting beta=1, the algorithm will perform no operation on the HDRI, there will be only a linear shrink of the dynamics (i.e. you are just making a gamma correction with gamma=1). Decreasing beta you will increase the effectiveness of the algorithm, in other words you will increase the compression of the dynamics making the details (and noise) much more evident. Low values of beta will create a very bright image with clear details and an unrealistic look, choosing a beta close to 1 you will get more realistic effects. The authors said to have used values between 0.8 and 0.9 with good results.
Color Saturation (Default 1)
As the name says, the color saturation. If the saturation equals 1, then you are not affecting the saturation at all: all the 3 RGB channels will be scaled with the same coefficients, so the original colors should be preserved. It can be seen like the "gamma" in the gamma-correction process, but it is applied to the three RGB components separately during scaling. That's why choosing 1 as saturation will cause no effect. If you choose saturation equal to zero, then you are telling the algorithm to discard all the information about color, it will use only the luminance of the image. The output can still have colors because of noise, but such colors might be completely different from the original. Using low values of saturation (like 0.2) will create a smooth grayscale effect.
Authors said to have used values between 0.4 and 0.6 with good results.
Pattanaik
pattanaik00 mult=1 cone=0.5 rod=0.5
When to use this algorithm
(Thanks to Ron Todd)
The key is to recognize that the ROD and CONE controls might work, to some extent, like the counterparts in our eyes: ROD controls color sensitivity while CONE controls light sensitivity.
My best results have been with high ROD value (0.75 to 0.99) and low CONE value (0.01 to 0.25). It seems that I can keep constant luminance if ROD + CONE = 1.00 or close but when CONE is too large the image becomes unsaturated even for large ROD values. The result is still very dependant on input HDR and on pre-gamma setting.
Parameters in detail
Reinhard 02
reinhard02 key=0.18 phi=1
When to use this algorithm
The pictures created with this algorithm will have an almost flat histogram, this means you are fully using the (limited) dynamic a standard 8-bit format can offer. The output will look realistic, keeping the details clear at any contrast.
Parameters in detail
Reinhard 04
reinhard04 brightness=-10 saturation=0.99
When to use this algorithm
It's a relatively simple algorithm with few processing on the original image. It will shrink the dynamics keeping the look very realistic. It does not do anything to stress the importance of details, to the general look will be quite blurred (you can limit this effect using a sharpening filter with Gimp or any other LDR program). It works better than other algorithms if the input image hasn't many details and it contains noise (e.g. night pictures taken with high ISO value)
Parameters in detail
The algorithm gives the possibility to change only two parameters whose meaning is clear, providing a good compromise between adaptability and easy-to-use.
The algorithm is basically this: each pixel whose value is "I" is replaced with a value given by I/(I+S), where "S" is a parameter that changes in each pixel and is computed according to the surrounding pixels. Smaller "S" will cause the pixel to be brighter, simulating a gamma correction with gamma>1.
Brightness (default -10)
The meaning is obvious, the greater the brightness, the brighter the image. From a mathematical point of view, increasing the brightness will decrease the average value of "S" in the previous relationship, thus increasing the overall brightness of all the pixels.
Saturation (default 0.99)
It obviously affects saturation. If it equals zero, the output image will be more similar to a grey-scale one (even if it will still have colours); increasing its value will give more saturated colours.
