Generation Process: Scoring

If automatic scoring is on (1), GenShade evaluates images of shaders in the ongoing population by comparing them to single or multiple target images. The multi-resolution image wavelet scoring metric is using the YIQ color system, with several unique considerations. To implement the notion of gender, each shader is scored in three ways: illumination only by considering the Y component of the YIQ system, chromaticity only by considering the I and Q components, and overall by considering all three components. To provide the capability of comparing generated images with multiple target, the most significant common wavelets among the multiple target images are compared to those of generated images. This allows matching the look of a group of related texture images, rather than simply matching a particular image of that texture.

The scoring system is using approximation and therefore it may not be 100% accurate, however it showed remarkable results in terms of directing resulting shaders towards target images of texture. Simply, some of the low scored shaders may be needed to introduce diversity and may contain valuable sub components.

The final results will be saved in the main or new MGP pools of each generation. In addition, you may browse through all generations to select desired shaders.

If interactive scoring is on, the MGP pools would not be updated every generation simply because there is no score. Shaders would be selected then from the initial MGP pools based on their rank. For example, a shader ranked 1st would have much more chance of getting selected than one that is ranked 50. To obtain the final results, the artist may browse through all generations to select desired shaders.

(1) please note the followings:

a. if you use a database of shaders that you have created, you will need to run the standalone tool decomposeimg to pre-calculate wavelets of database images.

b. turn the automatic scoring option on before loading a project or shader database.

If you encounter errors in this document, please email [email protected] with the URL, and enclose an example of the erroneous text, and what you feel to be the corrections, if possible.