brutpitt/glslsmartdenoise

Fast glsl deNoise spatial filter, with circular gaussian kernel, full configurable

Hi there 👋

glslsmartdenoise's Issues

Is this shader separable?

I noticed this shader is quite similar to a gaussian blur, which is a separable filter (blurring on just the X axis, and then blurring that image on just the Y axis gives the same results as doing a full convolution).

Do you think this could be turned into a separable version? I'm attempting that now, but I may have my math incorrect and I'm not 100% sure if it's even possible. If I come up with something I'll be sure to submit a PR, but I'm curious if you've already evaluated this.

Converted to exeldro/OBS-shaderfilter

I converted the shadertoy version of the shader to a shader filter for exeldro's obs-shaderfilter and it works great!

glslSmartDeNoise is now utilized in JWildfireSwan

Hi BrutPitt,
I just wanted to mention that I'm now using an adoption of your glslSmartDeNoise-algorithm in my project, which is an advanced flame-fractal-generator based on WebGl and Vaadin: https://github.com/thargor6/JWildfireSwan
There is also a live demo: https://herokuapp.overwhale.com/

Thanks for your work and best regards,
Andreas

Potential issues in the shader

Hello!

I've noticed a number of peculiar things in the fragment shader code, which I believe are bugs:

1. float invSigmaQx2PI = INV_PI * invSigmaQx2; // 1.0 / (sqrt(PI) * sigma)
   This line is not equal to 1.0 / (sqrt(PI) * sigma), which seems to simply be a comment error. However, this variable is also not used anywhere in the code, see point 3.

2. vec2 d = vec2(x,y)/size;
   Dividing d by the texture size implies that the standard deviation (sigma) is in normalized texture coordinates. In fact, dot(d, d) is so close to zero that the resulting denoised image looks identical if it is simply replaced with 0.0. I believe the correct value should be vec2(x, y);, without the division.

3. float blurFactor = exp( -dot(d , d) * invSigmaQx2 ) * invSigmaQx2;
   The multiplication by invSigmaQx2 seems incorrect and should be invSigmaQx2PI.

Thanks for the nice and simple denoising filter!

Edit:
Here's my hopefully fixed ShaderToy code, also switched to using texelFetch() instead of texture().
https://pastebin.com/FUScJxgK

Edit2:
Forcing float blurFactor = 1.0; gives visually indistuingishable results to the original shader, and most likely behaves more like a smart box filter than a smart gaussian blur. It does however still look good, and is presumably a decent bit cheaper to evaluate.

Using this shader with HDR tonemapping image: before or after?

I'm looking at integrating this denoising shader in Godot Engine and was wondering about denoising an HDR viewport. The engine uses a tonemapping shader to convert the viewport into a LDR image for display. For best results, should I run the denoising shader before or after tonemapping? My gut feeling tells me I should do this after tonemapping, but I'm not sure exactly.

This should be documented in the README and/or in the shader code. Thanks in advance for replying 🙂