Update with my findings here. From the test i started doing, it would appear that if you wanna go 4K with convolution Bloom, you’ll need a convolution Kernel with a higher resolution.
For instance back to the picture i posted earlier.
Results at commercial 4K (3840x2160) rendering for a single white/hot pixel Kernel:
For a kernel of resolution of:
2048px:
http://i.imgur.com/nbEa5j8m.jpg
4096px:
http://i.imgur.com/eLDTHIum.jpg
3072px:
http://i.imgur.com/hlPxSVCm.jpg
Every attempt below Kernel resolution of 3072px square ended up with corrupted results while you can see the actual luminance gradually going back to its 1080p-2440p value.
2560px:
http://i.imgur.com/wGTI8btm.jpg
2816px:
http://i.imgur.com/JKiJW0Lm.jpg
I stopped the dichotomy at 2816, so the range 2816-3072 hasn’t been evaluated, it’s more than enough to prove the point anyway.
You might wonder why i bothered with the resolution, the thing is that those texture files are huge.
Kernel resource size:
Default Engine Kernel at 2048px: 32768Kb
Single pixel 4K : 131072Kb
Single pixel 3K : 73728Kb
Using those results, i have created an ‘expanded’ default Engine Kernel of resolution 3072px, merely filling the extra pixel with black, and this is the result:
http://i.imgur.com/s2ophM0m.jpg
Again, as comparison, using the default engine Kernel of resolution 2048px, at 4K:
http://i.imgur.com/aFRAn5Um.jpg
And default engine Kernel at 1920x1080p:
http://i.imgur.com/YayAuVWm.jpg
In case you wanna try the convolution bloom at 4K , you can download the expanded kernel here: