This is the additional material accompanying the paper "Layered Weighted Blended Order-Independent Transparency"

The folders contain the performance plots, evaluations of the bin count, test images, and error images.

Due to upload restrictions we had to compress the images as jpgs. Please ignore any compression artifacts encountered. We tried to keep a high quality.

The performance plots are in the "performanceplots" folder. The file "perfplot_powerplant_auto.pdf" is a scaled version of the "perfplot_powerplant.pdf"

The images are in the "result images" folder divided into "result_images" which are the rendered images and "diff_images" which are the error images. We provide both a visulization of the root squared error (sqdiff) which clearly shows areas of very high errors and a logarithmically scaled version (logdiff) which gives a better overview of the error distribution.

The folder "frametime_vs_bincount" contains the rendering time plotted vs. the number of bins used. We can see a few interesting patterns in the data. In geometrically complex scenes, such as "dragons high poly" and "power plant", the linked list approach works extremely well and the limiting factor here is the rendering itself. On the other hand the layered rendering approach struggles to deliver good performance in these scenes as multiple rendering passes are required. The sudden "jumps" in the performance are due to the limit on the number of rendering targets that can be rendered during one pass.
Scenes with a low geometric complexity, such as "colored smoke" and "burning wood", show a clear linear increase in rendering time with respect to the number of bins which is desired as it shows that no bigger additional overhead is produced by our algorithm.

The folder "mse_vs_bincount" contains the number of bins plotted against the mean squared error (MSE). The scenes show generally the same exponential decrease in MSE with increasing bin count. Exceptions are scenes like "gray smoke" where the pure depth complexity of identically colored particles diminishes the benefit of having bins.
The slight shift in MSE between layered rendering and linked lists results from the fact that in the layered rendering approach 32-bins are rendered and smoothed in one pass and are then blended with with the framebuffer. Due to memory requirements we do not keep all bins in memory at the same time. Therefore, no smoothing between the bins of these batches is done in our implementation.

The file "fragstats.txt" contains the content of Table 1 plus an additional column with the average maximum per bin, i.e. for each pixel the bin with maximum number of fragments is counted and then averaged over all pixels.