High Pass Sharpening / Contrast Enhancement

Martigen wrote:

Ioxa wrote: Updated the Reshade + Framework Version.

-Changed the blurring technique used for sharpening. The new one has a much lower performance cost.

-Added options to curve the dark halos only or bright halos only. This article gives a pretty good overview on sharpening, halos, and why you may want to use stronger dark halos.

-Removed all the performance options and made the better performing settings the defaults.

Download for ReShade + Framework Version

I can understand removing the performance options to minimise complexity, but... there's quite a difference in the low-fi and high-fi version of the contrast mask, at least that I can see. It might be nice to keep this. And I presume the default options use 5 samples for each? I marginally noticed differences in quality with these disabled.

Also two questions:

1. The mid-tone versions are a great idea, I can see why you added them, especially on the contrast mask which tends to brighten everything. However, comparing screenshots using just 'mid-tone' and 'off', the mid-tone results seemed to apply *only* to brightened (and possibly dark, but hard to see there) areas. I expected mid-tone to be adjusted without brights being overly affected. Is this intentional, or did I misunderstand the purpose?

2. The CE mask width... help me understand how this should be used . At low levels (~1) it's clear the mask is, for want of a better phrase, low-accuracy. At the default of 3, it appears to look good, and doubling this to 6 seems to improve quality even further. But here's the rub: at 6 it shows what looks like a pretty accurate contrast mask, as some areas that are 'light' in 6 and representative of the source image are 'dark' at 3. I'm inclined to use 6 then, however given the end-result is blended and the goal is to create definition through contrast, the dark areas of '3' would actually -- I think, if understand this right -- create more definition. Is a higher value supposed to improve quality further, or does it get to a stage where higher values would be counter-productive?

That's all a bit wordy, so some images:

Source image (no effects):

MaskWidth = 3.0

MaskWidth = 6.0

Notice how, just for example in the three character portraits, when compared to the original image '3' is darkening the background of the portrait on the left, and on the right the male elf's face is dark whereas on '6' it is light, and closer to the source. You see it too in the final blended image but I'm undecided which looks better...

Settings used:

// ---- Contrast Enhancement Settings ----
#define CE_Mask               1 //[0|1|2|3] 0 = Off, 1 = Contrast Enhancement, 2 = Contrast De-enhancement?, 3 = Mid-tone Contrast Enhancement
#define CurveCE_Mask          1 //[0 or 1] 0 = Off, 1 = On.
#define CurveCE_Strength   0.20 //[0.00 - 1.00] Higher #'s = stronger effect.
#define CE_MaskWidth       3.00 //#'s > 1 = bigger radius, #'s < 1 = smaller radius
#define CE_MaskQuality        4 //[0 to 5] Higher #'s = more blur passes (Wider, smoother blur). Higher performance cost.
#define ViewCE_Mask           0 //Set to 1 to view the CE_Mask. Helpful when adjusting mask settings. 
#define CE_BlendMode          2 //1 = Soft Light(weak), 2 = Overlay(neutral), 3 = Multiply, 4 = Hard Light, 5 = Vivid Light, 6 = Soft Light#2(lighter), 7 = Soft Light#3(darker) 
#define CE_Strength        0.45 //[0.00 - 1.00]Strength of the Contrast Enhancement effect.

// ---- Performance and Misc Settings ---- (I recommend viewing the respective masks when adjusting these settings)
#define TextureFormat         0 //[0|1|2] 0 = RGBA8, 1 = RGBA16F(shows a little more detail than RGBA8), 2 = RGBA32F
#define SharpTexScale         0 //[0 or 1] 0 = Off, 1 = Scaled up texture for sharp mask. (Higher performance cost, slightly better quality.) 
#define SharpSamples          1 //[0 or 1] 0 = 3 samples, 1 = 5 samples. Setting to 0 slightly lowers performance cost and quality.
#define CETexScale            1 //[0 or 1] 0 = Scaled down texture for contrast enhancement mask (lower performance cost), 1 = No scaling. 
#define CE_Samples            1 //[0 or 1] 0 = 3 samples, 1 = 5 samples. Setting to 0 slightly lowers performance cost and quality.

Marty McFly wrote: Oh and I'm not sure that X passes with low radius gaussian blur is effective for big blur, you might consider increasing the gaussian tap count instead. Plus you can increase offset the more passes you have, i.e. if you go further than 1 pixel in first pass, you get artifacts but when you sample an already blurred image, you can shift the offsets by 10 pixels instead of 1 and you won't see any artifacts.

This is well above my head, but I did notice an improvement with more passes (which is why it's set to 4 above) but then I was also playing with the larger radius. Also thanks for all your work Marty!