SSAO Moving | Graphics Programming | Page 1

regal perch May 20, 2026, 3:47 PM

#

My ssao is having this weird issue up-close where the ao moves around with the camera.
Any ideas what i should check to figure out what could be causing this?

#

or it might just be an unavoidable issue from it being screen-space idk

shrewd jetty May 20, 2026, 4:07 PM

#

This looks many levels of wrong. What AO technique are you using?

regal perch May 20, 2026, 4:08 PM

#

i just sample random positions in a hemsiphere (using depth and normal textures)

#

wdym by "many levels of wrong" tho

#

as far as i can tell the only issue is the effect moving around when up-close (and the edges of the screen not having correct ao but thats to be expected)

shrewd jetty May 20, 2026, 4:19 PM

#

Alright, so imma warn you against doing that. It’s simple but looks bad, figure out horizons.

How do you reconstruct position from depth?

regal perch May 20, 2026, 4:19 PM

#

shrewd jetty Alright, so imma warn you against doing that. It’s simple but looks bad, figure ...

i do this: (i have reversed infinite far-plane depth)

vec3 reconstructViewPos(vec2 uvPos, float depth) {
    vec4 clip = vec4((uvPos*2)-1, depth, 1.f);
    vec4 view = inverse(globalUbo.proj)*clip;
    return view.xyz/view.w;
}

#

also i have heard of horizon based ssao but it doesnt look any better to me

#

maybe im just not looking at the right comparisons tho

shrewd jetty May 20, 2026, 4:20 PM

#

It looks much better. I don’t think your average joe can’t tell, though.

#

Hmm, right. In my time I’ve just did clip, depth remap and multiply xy etc etc, without inverting stuff.

#

I’d try that

#

Gimme a sec, imma get you a snippet

regal perch May 20, 2026, 4:22 PM

#

shrewd jetty Hmm, right. In my time I’ve just did clip, depth remap and multiply xy etc etc, ...

i mean i'd think its important to invert the projection matrix when using it to undo a previous projection

shrewd jetty May 20, 2026, 4:23 PM

#

Yeah, it’s just perspective though, correct?

regal perch May 20, 2026, 4:23 PM

#

yeah

shrewd jetty May 20, 2026, 4:23 PM

#

Can avoid that

regal perch May 20, 2026, 4:23 PM

#

oh are you meaning like dont use the projection matrix at all?

shrewd jetty May 20, 2026, 4:23 PM

#

Just as a test

regal perch May 20, 2026, 4:25 PM

#

hm if i dont use the projection matrix than there just no ao appearing at all

shrewd jetty May 20, 2026, 4:25 PM

#

float3 getViewPos(float2 uv, float z) {
    float fin_z = remap(z, 0., 1., n, f);
    float2 norm = (uv - 0.5.xx) * 2.0;
    float fl = 2.0 * tan(FOV * 0.5);
    
    float3 pos = float3(norm * fl * fin_z, fin_z);
    pos.y *= rcp(BUFFER_ASPECT_RATIO);
    
    return pos;
}```

n is nearplane, f is farplane, fl is focal length

#

you'll need to reverse it, though

#

i seem to have misplaced the inversion, but you can figure it out yourself

regal perch May 20, 2026, 4:26 PM

#

ngl ive got no idea what that is doin

#

i dont think anything is wrong with my current reconstruction tho

shrewd jetty May 20, 2026, 4:27 PM

#

well, okay, what rng are you using

regal perch May 20, 2026, 4:28 PM

#

i use the randomVec for offset and randomFloat for radius (which now that i think about it is redundant)

const vec3 SSAO_NOISE[16] = vec3[](
    vec3( 0.7071,  0.7071, 0.0), vec3(-0.3827,  0.9239, 0.0), vec3( 0.9808, -0.1951, 0.0), vec3(-0.8315, -0.5556, 0.0),
    vec3( 0.1951, -0.9808, 0.0), vec3( 0.9239,  0.3827, 0.0), vec3(-0.5556,  0.8315, 0.0), vec3(-0.9808,  0.1951, 0.0),
    vec3( 0.3827, -0.9239, 0.0), vec3( 0.8315,  0.5556, 0.0), vec3(-0.7071, -0.7071, 0.0), vec3(-0.1951,  0.9808, 0.0),
    vec3( 0.5556, -0.8315, 0.0), vec3( 0.9808,  0.1951, 0.0), vec3(-0.3827, -0.9239, 0.0), vec3(-0.9239, -0.3827, 0.0));
vec3 randomVec() {
    int x = int(mod(gl_FragCoord.x, 4.0));
    int y = int(mod(gl_FragCoord.y, 4.0));
    return SSAO_NOISE[x * 4 + y];
}
float randomFloat(int y) {
    int x = int(mod(gl_FragCoord.x, 4.0));
    return SSAO_NOISE[x * 4 + y].x;
}

shrewd jetty May 20, 2026, 4:28 PM

#

boi

regal perch May 20, 2026, 4:28 PM

#

?

shrewd jetty May 20, 2026, 4:28 PM

#

just use this for now

#

https://www.shadertoy.com/view/4djSRW

#

because its quite likely you're looking at correlations that appear because of

#

well

#

mod()

#

on coords

#

oh, and also, random inside sphere, right

#

hmm

regal perch May 20, 2026, 4:33 PM

#

shrewd jetty oh, and also, random inside sphere, right

wdym

#

ok so i get a random vector to use as the direction

#

multiply it by the random float used as radius

#

and i use the normal to make sure it is a hemisphere

shrewd jetty May 20, 2026, 4:34 PM

#

random vector is not the same as random direction

regal perch May 20, 2026, 4:35 PM

#

how so?

shrewd jetty May 20, 2026, 4:35 PM

#

imagine a cube

#

that's where the random vector lies

#

now smooth it out

#

you'll be cutting away at the corners, while the faces will sort of remain remain

#

that's my best intuitive explanation

#

also, not what you need

regal perch May 20, 2026, 4:36 PM

#

yeah idk what you mean by any of that

shrewd jetty May 20, 2026, 4:36 PM

#

also, you're not picking a random direction

#

unless you're not doing crytek ao

regal perch May 20, 2026, 4:37 PM

#

idk what crytek is so i probably aint doin that

shrewd jetty May 20, 2026, 4:37 PM

#

you're picking a random point in the volume

shrewd jetty May 20, 2026, 4:37 PM

#

regal perch idk what crytek is so i probably aint doin that

where you got this general impl from?

regal perch May 20, 2026, 4:38 PM

#

tbh i dont remember, i made this like a month ago

#

i think i used multiple guides tho

shrewd jetty May 20, 2026, 4:39 PM

#

rn it seems to me youre trying to make de-assified crytek ao

#

which is the simplest

regal perch May 20, 2026, 4:39 PM

#

shrewd jetty because its quite likely you're looking at correlations that appear because of

this is what my randomVec looks like visualized

#

i do see a pattern

#

so yeah its probably wrong

shrewd jetty May 20, 2026, 4:40 PM

#

yeah it's bad

#

ok, use the hashes from the link

regal perch May 20, 2026, 4:40 PM

#

ok

shrewd jetty May 20, 2026, 4:40 PM

#

only for testing, replace with something else later

#

do AO debug view, output just the AO

regal perch May 20, 2026, 4:42 PM

#

im confused about the hashing

#

one of the methods wants a vec2 and outputs just one float

#

the other wants a vec3 and outputs a vec3

#

i do want a vec3 but idk what i would input for the third number

#

depth maybe?

#

they are using time in the example

#

but i dont think i want the ao moving over time

shrewd jetty May 20, 2026, 4:44 PM

#

not depth

#

it can be constant

#

just 1.0

regal perch May 20, 2026, 4:44 PM

#

ah

#

yeah this looks much more random

shrewd jetty May 20, 2026, 4:46 PM

#

discord agrees

#

use that for now

#

also AO buffer!!

regal perch May 20, 2026, 4:46 PM

#

alright

#

#

looks about the same to me

#

still has the moving around problem up-close too

shrewd jetty May 20, 2026, 4:49 PM

#

hm

#

so first off, you want points in a sphere

#

a simple (but not efficient on GPU) method is to just reroll the RNG until length < 1

regal perch May 20, 2026, 4:53 PM

#

shrewd jetty a simple (but not efficient on GPU) method is to just reroll the RNG until lengt...

the rng is always less than 1 i think

#

and why would i want the points in a sphere instead of hemisphere?

regal perch May 20, 2026, 4:55 PM

#

regal perch the rng is always less than 1 i think

yeah the rng never makes a value greater than 1

#

i made it show red if any channel is greater than 1 and green if its not

regal perch May 20, 2026, 4:57 PM

#

shrewd jetty

also not sure what your tryna show here

shrewd jetty May 20, 2026, 5:00 PM

#

Length(rng)

#

Or better yet, dot(rng,rng) for that one extra bit of speed

regal perch May 20, 2026, 5:01 PM

#

for what

#

idk where you want me to put length(rng)

shrewd jetty May 20, 2026, 5:02 PM

#

Ok, what I’m saying is that your random numbers need to be in a sphere with radius one in accordance with crytek ao

#

Just because each individual component of a vector is < 1, doesn’t mean the vectors length has to be < 1

regal perch May 20, 2026, 5:03 PM

#

ohhh

shrewd jetty May 20, 2026, 5:03 PM

#

Imagine a vector (0,1,1)

#

That’s length sqrt(2), a bit over 1.41

regal perch May 20, 2026, 5:03 PM

#

why does the length matter tho i dont currently get the length anywhere

shrewd jetty May 20, 2026, 5:03 PM

#

Ok so at the corner of the block, right?

#

There’s a sharp 90 degree angle

#

That should dissapear

#

(I think at least)

regal perch May 20, 2026, 5:04 PM

#

ohhh i see

shrewd jetty May 20, 2026, 5:04 PM

#

But overall, that’s what the technique requires

regal perch May 20, 2026, 5:05 PM

#

couldnt i just do hash = hash/length(hash)

#

to avoid the >1 length
apparently that doesnt work like i thought it would, still has some greater than 1 lengths

shrewd jetty May 20, 2026, 5:07 PM

#

shrewd jetty imagine a cube

Basically this

#

You can, but your samples will just be non-uniformly distributed on the surface of a sphere

#

They need to be (at least for the most part), uniform and inside the sphere

regal perch May 20, 2026, 5:08 PM

#

i could just divide the result by 1.5 or smth

#

whatever the max possible length is (1,1,1)

#

1.732

shrewd jetty May 20, 2026, 5:09 PM

#

What is it with people and bandaiding shit

#

Yes, you can

regal perch May 20, 2026, 5:09 PM

#

bandaid?

shrewd jetty May 20, 2026, 5:09 PM

#

However, you can a smaller cube

regal perch May 20, 2026, 5:09 PM

#

its reducing it to the right range

shrewd jetty May 20, 2026, 5:09 PM

#

Which is, for the exact same reasons, not what you want

regal perch May 20, 2026, 5:09 PM

#

its not?

#

i dont understand why

shrewd jetty May 20, 2026, 5:10 PM

#

https://karthikkaranth.me/blog/generating-random-points-in-a-sphere/

Generating Random Points in a Sphere | Karthik Karanth

Visualization of different methods of generating random points inside a sphere

#

You need this.

#

As for why, see if drawing it can help

#

Same shit in 2D

#

Or just plt that shit, because numeric experiments are based

regal perch May 20, 2026, 5:11 PM

#

plt?

shrewd jetty May 20, 2026, 5:12 PM

#

Matplotlib

#

Python library that lets you easily plot data

regal perch May 20, 2026, 5:12 PM

#

i mean the dividing by 1.732 seems to work

shrewd jetty May 20, 2026, 5:12 PM

#

You can just, run that calculation and visually see the clumping near the corners

regal perch May 20, 2026, 5:12 PM

#

although none of this affects the up-close moving around issue

#

also smth i noticed is the hashing made the ao noisy

shrewd jetty May 20, 2026, 5:13 PM

#

regal perch i mean the dividing by 1.732 seems to work

It’s different sorts of wrong though.

shrewd jetty May 20, 2026, 5:13 PM

#

regal perch also smth i noticed is the hashing made the ao noisy

And yeah, when you add noise you get noise.

#

But before, you had 16 or so predetermined values which caused banding instead

regal perch May 20, 2026, 5:13 PM

#

ohh

#

i still have the banding

#

its just also noisy

#

shrewd jetty May 20, 2026, 5:14 PM

#

I see no bands here tbh

#

Oh nvm

#

You’re not slicing the array with those values, right?

#

Just using them directly?

regal perch May 20, 2026, 5:14 PM

#

im just using them directly

shrewd jetty May 20, 2026, 5:15 PM

#

Guh

regal perch May 20, 2026, 5:15 PM

#

maybe the kernel is causing it idk

shrewd jetty May 20, 2026, 5:15 PM

#

Probably just looking at the wrong distribution

#

Yes, the kernel is wrong

regal perch May 20, 2026, 5:16 PM

#

vec3 SSAO_KERNEL[32] = vec3[](vec3( 0.021,  0.183,  0.512), vec3( 0.392,  0.041,  0.734), vec3(-0.221,  0.114,  0.612), vec3( 0.134, -0.287,  0.553), vec3(-0.341, -0.102,  0.487), vec3( 0.287,  0.331,  0.682), vec3(-0.129,  0.412,  0.731), vec3( 0.512, -0.221,  0.612), vec3(-0.412, -0.331,  0.682), vec3( 0.221,  0.129,  0.341), vec3(-0.183, -0.021,  0.512), vec3( 0.331, -0.412,  0.731), vec3(-0.041,  0.392,  0.734), vec3( 0.102, -0.341,  0.487), vec3(-0.287,  0.134,  0.553), vec3( 0.412,  0.287,  0.612), vec3(-0.512, -0.183,  0.512), vec3( 0.341, -0.102,  0.487), vec3(-0.129,  0.221,  0.341), vec3( 0.183,  0.412,  0.731), vec3(-0.392,  0.041,  0.734), vec3( 0.102, -0.512,  0.612), vec3(-0.221, -0.392,  0.682), vec3( 0.331,  0.129,  0.341), vec3(-0.041, -0.183,  0.512), vec3( 0.287, -0.412,  0.731), vec3(-0.134,  0.341,  0.487), vec3( 0.412, -0.287,  0.612), vec3(-0.512,  0.183,  0.512), vec3( 0.341,  0.102,  0.487), vec3(-0.129, -0.221,  0.341), vec3( 0.183, -0.412,  0.731));
for (int i = 0; i < KERNEL_SIZE; i++) {
  vec3 sampleVec = TBN * SSAO_KERNEL[i];

hmm

shrewd jetty May 20, 2026, 5:16 PM

#

Ok so

regal perch May 20, 2026, 5:16 PM

#

maybe i should use the random number instead of a kernel?

shrewd jetty May 20, 2026, 5:17 PM

#

If you have tabulated values that are guaranteed correct inside of a sphere, you can roll with those

#

Though I’m a stickler for actually generating white noise until you run into quality issues with it

#

But when shit doesn’t look correct, getting it to the most minimal point is the right move

regal perch May 20, 2026, 5:18 PM

#

yeah using a random number instead of a kernel got rid of the banding

#

its even more noisy now tho

#

and less stable

shrewd jetty May 20, 2026, 5:18 PM

#

Yes! But more debugfavle

#

So, first off, the hash does fract()

#

So it returns a value from 0-1

regal perch May 20, 2026, 5:19 PM

#

right

shrewd jetty May 20, 2026, 5:19 PM

#

So just -0.5.xxx that shit

regal perch May 20, 2026, 5:19 PM

#

oh yeah true i forgot i need -1 to 1 not 0 to 1

shrewd jetty May 20, 2026, 5:20 PM

#

It just needs to be symmetrical around 0, you can skip the *2, that’s just radius for now

regal perch May 20, 2026, 5:20 PM

#

ooh that actually looks super good if you ignore the noise

#

theres highlights on edges

shrewd jetty May 20, 2026, 5:21 PM

#

Which is very wrong

regal perch May 20, 2026, 5:21 PM

#

but looks good

shrewd jetty May 20, 2026, 5:21 PM

#

But, the hemisphere check is supposed to fix that

regal perch May 20, 2026, 5:21 PM

#

yeah odd that it isnt

shrewd jetty May 20, 2026, 5:21 PM

#

regal perch but looks good

Debatable, I know a lot of people will fry you for that

#

I like it myself

#

Though there are probably cheaper ways to achieve it

#

Also, I think it’s still asymmetrical

#

You remapped all components of the vector, right?

regal perch May 20, 2026, 5:23 PM

#

asymmetry would explain the issue with the moving around up close

#

wdym by remapped

shrewd jetty May 20, 2026, 5:23 PM

#

As in -.5

#

You need to do that for every component

#

Or just -0.5.xxx the entire thing

regal perch May 20, 2026, 5:23 PM

#

yeah:

vec3 hash(){
    vec3 p = vec3(uv.xy, 1);
    p = vec3(dot(p, vec3(127.1,311.7, 74.7)), dot(p, vec3(269.5,183.3,246.1)), dot(p, vec3(113.5,271.9,124.6)));
    return (fract(sin(p)*43758.5453123)-0.5f)*2;
}

shrewd jetty May 20, 2026, 5:24 PM

#

Add the .xxx swizzle so that I may sleep tonight

regal perch May 20, 2026, 5:24 PM

#

lol ok

shrewd jetty May 20, 2026, 5:24 PM

#

Idk the default behavior for glsl

regal perch May 20, 2026, 5:24 PM

#

it multiplies all components

#

but im confused on what you mean by .xxx

#

where do i put it, never seen that before

shrewd jetty May 20, 2026, 5:25 PM

#

Yeah, though scalar + vector is technically undefined in a mathematical sense

shrewd jetty May 20, 2026, 5:25 PM

#

regal perch where do i put it, never seen that before

After .5

regal perch May 20, 2026, 5:25 PM

#

ok i put the .xxx

#

didnt change anything visually

#

i'll just keep it there tho

shrewd jetty May 20, 2026, 5:26 PM

#

Ok, good. Explicit is better then implicit (to a point bleakekw )

regal perch May 20, 2026, 5:26 PM

#

looks like it is asymmetry yeah

shrewd jetty May 20, 2026, 5:26 PM

#

Yeah

#

So

#

That’s the box random numbers showing themselves

shrewd jetty May 20, 2026, 5:27 PM

#

shrewd jetty https://karthikkaranth.me/blog/generating-random-points-in-a-sphere/

This is the fix.

regal perch May 20, 2026, 5:27 PM

#

which one of the solutions there should i do

#

i could just do the last one it looks pretty uniform

shrewd jetty May 20, 2026, 5:28 PM

#

Whatever is the simplest

#

Last one is good

#

Avoid the first one, its probably really bad for GPU or something

regal perch May 20, 2026, 5:28 PM

#

oof the last one isnt possible

#

it uses Math.random

#

i think they all do actually

shrewd jetty May 20, 2026, 5:29 PM

#

Fam

regal perch May 20, 2026, 5:29 PM

#

maybe i could re-use the hashing for that instead

shrewd jetty May 20, 2026, 5:29 PM

#

https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects/Math#static_methods

MDN Web Docs

Math - JavaScript | MDN

The Math namespace object contains static properties and methods for mathematical constants and functions.

regal perch May 20, 2026, 5:29 PM

#

shrewd jetty https://developer.mozilla.org/en-US/docs/Web/JavaScript/Reference/Global_Objects...

well theres no javascript in glsl

shrewd jetty May 20, 2026, 5:29 PM

#

Yeah

#

So what you need is the information on where the numbers are

#

They’re probably 0-1

#

So you can just reuse the hash

regal perch May 20, 2026, 5:30 PM

#

shrewd jetty So what you need is the information on where the numbers are

what are you talkin about?

shrewd jetty May 20, 2026, 5:30 PM

#

Remove the .5.xxx and *2 though

shrewd jetty May 20, 2026, 5:30 PM

#

regal perch what are you talkin about?

The examples are written in js. This is the documentation for js’s Math.

regal perch May 20, 2026, 5:31 PM

#

where what numbers are tho

shrewd jetty May 20, 2026, 5:31 PM

#

So scroll down (or don’t, I already told you what to do)

#

And look at the range

regal perch May 20, 2026, 5:31 PM

#

range of what??

shrewd jetty May 20, 2026, 5:31 PM

#

The rng

regal perch May 20, 2026, 5:31 PM

#

idk what you are talking about

#

i cant use javascript methods in glsl

shrewd jetty May 20, 2026, 5:32 PM

#

Yes, you use a hash

regal perch May 20, 2026, 5:32 PM

#

right

shrewd jetty May 20, 2026, 5:32 PM

#

But you need to match up the example code and GLSL

regal perch May 20, 2026, 5:32 PM

#

ohhh

#

yeah i see what you mean

shrewd jetty May 20, 2026, 5:32 PM

#

And just to be sure, you should check what Math.random() does

regal perch May 20, 2026, 5:32 PM

#

i need to figure out what range javascript's Math.random() returns

shrewd jetty May 20, 2026, 5:32 PM

#

Genau

regal perch May 20, 2026, 5:32 PM

#

yeah its the same as the hash perfect

shrewd jetty May 20, 2026, 5:32 PM

#

Yep

#

Don’t forget you’ve modified it

regal perch May 20, 2026, 5:33 PM

#

yeah i undid the +0.5 and /2 thing

#

also do you know what the x:x, y:y, z:z from the example means?

#

is that some kind of bitwise operation

#

idk

#

x:x aint valid in glsl

#

maybe its just a weird way to make a vec3 in some other language idk

shrewd jetty May 20, 2026, 5:36 PM

#

That’s iirc just shorthand init of a class that contains x,y,z

regal perch May 20, 2026, 5:36 PM

#

ah

#

and ofc glsl has no method to get the cube root

shrewd jetty May 20, 2026, 5:37 PM

#

Heard of jsons? Ofc you did, you probably had the fun of fucking around with them because modding.

regal perch May 20, 2026, 5:37 PM

#

yeah

shrewd jetty May 20, 2026, 5:37 PM

#

regal perch and ofc glsl has no method to get the cube root

Use pow

regal perch May 20, 2026, 5:37 PM

#

arent those two different operations?

#

i could look into how cubic roots are calculated and recreate that method maybe

shrewd jetty May 20, 2026, 5:38 PM

#

Elementary school level math

regal perch May 20, 2026, 5:38 PM

#

bro they dont teach cubic roots at any stage of education

#

maybe in college actually

#

but yeah not in normal schools

shrewd jetty May 20, 2026, 5:39 PM

#

A root is just a non-whole power

#

Just uhh

#

Pow(…,1./3.)

#

Wonder if that has edge cases in negative numbers…

regal perch May 20, 2026, 5:40 PM

#

i googled it and it said smth like that

#

also need to abs it if theres negatives

shrewd jetty May 20, 2026, 5:41 PM

#

Ah, it’s UB

regal perch May 20, 2026, 5:41 PM

#

but i aint got negatives from the hash anyways

#

so should be able to just do pow(hash.z, 0.333333333f);

shrewd jetty May 20, 2026, 5:41 PM

#

regal perch also need to abs it if theres negatives

You can, but in this case since you’d a cube root it’s sign(…)*pow(…, 1./3.)

regal perch May 20, 2026, 5:42 PM

#

whats the point in sign if its always 1

shrewd jetty May 20, 2026, 5:42 PM

#

You’d

#

Not should, mb

regal perch May 20, 2026, 5:42 PM

#

ah

shrewd jetty May 20, 2026, 5:42 PM

#

Yep, all nice and defined here

regal perch May 20, 2026, 5:43 PM

#

implementing the last example from that website doesnt change the results noticably

#

maybe the issue is that randomVec is the same every time i call it from one pixel

shrewd jetty May 20, 2026, 5:43 PM

#

Yeah, it is

#

But show pls

regal perch May 20, 2026, 5:44 PM

#

float getAO(float depth, vec3 normal) {
    vec3 posVS = reconstructViewPos(uv.xy, depth);
    vec3 normalVS = normalize((globalUbo.view * vec4(normal.xyz, 0.f)).xyz);
    vec3 randVec = randomVec();
    vec3 tangent = normalize(randVec - normalVS * dot(randVec, normalVS));
    vec3 bitangent = cross(normalVS, tangent);
    mat3 TBN = mat3(tangent, bitangent, normalVS);
    float occlusion = 0.f;
    for (int i = 0; i < KERNEL_SIZE; i++) {
        vec3 sampleVec = TBN * randomVec();//SSAO_KERNEL[i];

shrewd jetty May 20, 2026, 5:44 PM

#

ok

#

sampleVec, where is it used?

#

Also, show debug pic

regal perch May 20, 2026, 5:45 PM

#

sampleVec = posVS + sampleVec * radius;
vec4 offset = globalUbo.proj * vec4(sampleVec, 1.0);
offset.xyz /= offset.w;
vec2 sampleUV = (offset.xy*0.5)+0.5;
if (!(sampleUV.x < 0 || sampleUV.x > 1 || sampleUV.y < 0 || sampleUV.y > 1)) {
    vec3 sampleVS = reconstructViewPos(sampleUV, texture(ddaDepth, sampleUV).r);
    float rangeCheck = smoothstep(0.f, 1.f, (radius/AO_STRENGTH)/length(sampleVS-posVS));
    vec3 viewDirVS = normalize(posVS);
    bool occluded = sampleVS.z <= sampleVec.z-0.05f;
    occlusion += (occluded ? 1.f : 0.f) * rangeCheck * AO_STRENGTH;
}

#

radius is just 1

shrewd jetty May 20, 2026, 5:46 PM

#

Any reason why you TBN* it?

regal perch May 20, 2026, 5:46 PM

#

for the hemisphere

shrewd jetty May 20, 2026, 5:46 PM

#

Don’t for now

#

Also thats wrong I believe

#

Plus you said you liked the outlines 😛

regal perch May 20, 2026, 5:47 PM

#

if i dont tbn it then it gets very broken

shrewd jetty May 20, 2026, 5:47 PM

#

So it’s not hemisphere then

regal perch May 20, 2026, 5:47 PM

#

yeah i get the outlines with the tbn lol

#

well it was hemisphere

#

with the old random method

regal perch May 20, 2026, 5:48 PM

#

regal perch if i dont tbn it then it gets very broken

shrewd jetty May 20, 2026, 5:48 PM

#

That’s just because the points were distributed as such

shrewd jetty May 20, 2026, 5:48 PM

#

regal perch

Oh fuck

#

Radius

#

Has to be the radius, riight?

regal perch May 20, 2026, 5:48 PM

#

radius is just 1

regal perch May 20, 2026, 5:48 PM

#

regal perch

wait disregard that

#

it wasnt caused by removing the tbn

#

it was from me making the randomvec inputs different in the loop

#

this is without tbn

#

and yeah its perfectly stable now

#

no asymmetry

shrewd jetty May 20, 2026, 5:50 PM

#

Looks correct

#

Now, the original “motion” issue

regal perch May 20, 2026, 5:50 PM

#

the asymmetry was the motion issue

#

its gone now

shrewd jetty May 20, 2026, 5:50 PM

#

Neat!

regal perch May 20, 2026, 5:50 PM

#

and the blur should get rid of the noise

#

once i add blur

#

so i think its working good

shrewd jetty May 20, 2026, 5:50 PM

#

Well, it’s a lot of noise

regal perch May 20, 2026, 5:51 PM

#

true

shrewd jetty May 20, 2026, 5:51 PM

#

The best way to get rid of noise is to not create it

regal perch May 20, 2026, 5:51 PM

#

ah like by going back to the kernel? (nvm that breaks everything)

shrewd jetty May 20, 2026, 5:51 PM

#

In other words, steal a GTAO impl

#

It’s lower integration dimensionality will make your life way easier

regal perch May 20, 2026, 5:51 PM

#

i'll google what that is

#

oof is it actually 3-6x more expensive than ssao?

shrewd jetty May 20, 2026, 5:54 PM

#

… as for noise reduction

Don’t blur, use a basic À-Trous with a few passes
Temporal accumulation is really good for quality, but is difficult to get perfect
avoid white noise and hashes. they’re really good to debug, but ass for production. See blue noise. It’s also easier to do so with GTAO.

shrewd jetty May 20, 2026, 5:54 PM

#

regal perch oof is it actually 3-6x more expensive than ssao?

Sort of? It is more expensive ALU wise. But it is much higher quality. So you can do less samples. And get better denoising.

#

Pretty much, it’s worth it, imo

#

I’ll also tell you to avoid bitmasking for the time being, it’s cool but much more difficult and is missing the ground truth of GTAO if used as-is

regal perch May 20, 2026, 5:56 PM

#

i dont think im doin any bitmasking

shrewd jetty May 20, 2026, 5:57 PM

#

Yeah

#

Stick with barebones GTAO

regal perch May 20, 2026, 5:57 PM

#

is it a good idea to change the randomVec in my loop to be different from the one outside of it?
randomVec(uv.xy+(i/100))

#

like by using the iteration

shrewd jetty May 20, 2026, 5:59 PM

#

Every time you need a new random number it should be unique and ideally uncorrelated to an extent. That’s why .nextRandom() is a thing in rng patterns.

#

Don’t uv+something, use an rng with extra inputs

regal perch May 20, 2026, 5:59 PM

#

what inputs tho

shrewd jetty May 20, 2026, 5:59 PM

#

UV, index, optionally time

#

3 or 4

regal perch May 20, 2026, 6:00 PM

#

index is what i was just using tho

shrewd jetty May 20, 2026, 6:00 PM

#

ah

#

What’s the issue then?

regal perch May 20, 2026, 6:00 PM

#

or wait i see what you mean

#

i shouldnt add the index to the uv

#

i should make the 3rd parameter in the vec3 the hash uses

shrewd jetty May 20, 2026, 6:00 PM

#

Exactly

#

Should once again avoid correlations, which is usually good

#

Anyhow, imma log off for now. You should probably shoot me a dm or something if you need me.

regal perch May 20, 2026, 6:02 PM

#

alright

#

thanks for the help

regal perch May 20, 2026, 7:16 PM

#

i cant find anything good for how to implement gtao

#

so i think maybe i'll just add taa and hope it hides the noise

#

(also tried using blue noise but it brought back banding and asymetry / ao moving around up-close)

shrewd jetty May 20, 2026, 7:28 PM

#

regal perch i cant find anything good for how to implement gtao

The paper has pseudocode and looking at an impl will do you about as well. Otherwise read up on (genuinely) SSR and HBAO. Beyond that if you can grasp the math, that’s great. Otherwise, just kinda debug issues with it until it works.

#SSAO Moving