Perusing various articles on different blur algorithms for a bloom step, they all seem to regard a gaussian blur as the convolution of choice. However, after putting it to the test with a large array of conditions, I came to the conclusion that a gaussian blur isn't as visually appealing as one would expect, not even taking artifacts incurred by cheap approximations into account. No matter how large a kernel and standard deviation I use, the bloom cuts off way too abruptly in a circle around bright centers.

By artificially deviating from a standard gaussian curve and adding more weight to the center of a kernel without reducing the tail too much, I achieve results that I find to be more pleasing and in line with what I expect bloom to look like. Now, arbitrarily tuning the weights with no physical or mathematical basis isn't exactly ideal, so I am wondering if there is an appropriate alternative to the good old bell curve for this purpose.

Left-right-up-down filtering with IIR filters can approximate Gaussian filters with a very long (full image) tail very well. IIRC, Microsoft Research had a paper on how to do this efficiently on GPUs a few years ago.

EDIT: Found it: GPU-Efficient Recursive Filtering and Summed-Area Tables by Nehab et al.: http://hhoppe.com/recursive.pdf 

The abrupt circle cutoff is probably exactly due to the cheap approximations. A true Gaussian filter has an infinite support

48 minutes ago, l0calh05t said:

The abrupt circle cutoff is probably exactly due to the cheap approximations. A true Gaussian filter has an infinite support

Distant contributions quickly become negligible with anything but screen-wide blurs, and you have yourself a myopia-simulator by then.

It's not strictly speaking abrupt, but it's obvious that you are looking at a moderately concentrated disk, whereas real bright lights seen by your eyes tend to subtlety wash out a better part of your entire field of view and only bleed strongly within a degree of the light source.

53 minutes ago, silikone said:

Distant contributions quickly become negligible with anything but screen-wide blurs

Are you sure about the "negligible" part? Real HDR values can be many orders of magnitude apart, making what is negligible in LDR significant.

17 hours ago, l0calh05t said:

Are you sure about the "negligible" part? Real HDR values can be many orders of magnitude apart, making what is negligible in LDR significant.

Even at max FP16 values of ~65K, a gaussian kernel with a sigma of 1 modulates such a value to levels below the standard RGB range at any pixel radius consisting of two digits.

Of course, that's not a very blurry kernel (and you probably wouldn't want to use full-res pixels), but the higher you go, the less energy is concentrated near the center of a kernel, weakening the bleeding glow effect seen immediately around bright objects.

In essence, you want to retain both the soft round shape of a high-radius kernel and the high peak of a low-radius kernel.

Here's a stock photo that demonstrates a mixture of soft and hard glows.

One distribution you could look into that might match your expectations better is the Cauchy distribution (https://en.wikipedia.org/wiki/Cauchy_distribution) it has a sharp peak and a very wide, low intensity tail. As the Gaussian is non-physical anyway, there's nothing wrong with trying different distributions.

2 hours ago, silikone said:

Even at max FP16 values of ~65K, a gaussian kernel with a sigma of 1 modulates such a value to levels below the standard RGB range at any pixel radius consisting of two digits.

One more thought: are you correctly applying gamma or outputting the clamped result of the convolution as an SDR image as-is?

Just now, l0calh05t said:

One distribution you could look into that might match your expectations better is the Cauchy distribution (https://en.wikipedia.org/wiki/Cauchy_distribution) it has a sharp peak and a very wide, low intensity tail. As the Gaussian is non-physical anyway, there's nothing wrong with trying different distributions.

One more thought: are you correctly applying gamma or outputting the clamped result of the convolution as an SDR image as-is?

Thanks, that seems to be a candidate for what I was looking for. I don't mind non-physical models as long as they are backed up by well-established maths. I am applying sRGB gamma correction, for the record.