Tài

I make stuff

[Tài @LeLocTai]

@jamesonhodge Is this the same as pink trombone or did you add more terms?

[Tài @LeLocTai]

@SebAaltonen @lisyarus Ahh, i guess it is the ggx d term then? Saw it when reading about google filament once. I assume it's not point light because you'd use cem yuksel non singular style light.

[Tài @LeLocTai]

@lisyarus @SebAaltonen Aliasing when downsampling for bloom

[Tài @LeLocTai]

@charles_grassi You are giving the model mip samples per steps & a few toward the suns, so this is essentially cone tracing with an mlp post processor. That's what you should have compared with for equal time/quality comparison. This is apples and oranges.

[Tài @LeLocTai]

@thinkrectangles High quality *and* fast blur is actually very hard. Apple is the only one who do it right. Samsung have no delay but heavily undersample so it look "noisy" at low blur. Chrome and Firefox is bad too. But looking at this thread I guess they're good enough that most don't care

[Tài @LeLocTai]

@XorDev 0->10. Dsp

[Tài @LeLocTai]

@SebAaltonen @LundukeJournal I care even if I don't use it. It'd put pressure on the duopoly. Or monopoly if you don't have the money for a mac. Right now it is not, which i why MS can get away with so much anti consumer behaviours. Good thing Valve is changing it

[Tài @LeLocTai]

Took a break from some deep algorithm works to implement some basic but much needed features for Paraform

[Tài @LeLocTai]

@vanilagy Refraction aliasing is very hard to deal with. AA is averaging the content of 1px. On tight bends, 1px can compress the content of a very large area. Worse, this content differ greatly even for neighbouring px, so you can't use prefiltering like mipmap. It's a perf trade off

[Tài @LeLocTai]

@KostasAAA Gaussian variance is incorrect after normalization and also especially when sample count <=3. This matter for radius continuity, because variance from the deepest level is greatly magnified 4/4

[Tài @LeLocTai]

@KostasAAA Using 2 wide box for downsampling is cheap but can alias severely. Wider downsampling kernel is much better but also much more expensive. It is a bit compensated as you need less passes on average, but this is mostly for when you care about temporal stability 3/n

[Tài @LeLocTai]

Gaussian variance is incorrect after normalization and also especially when sample count <=3. This matter for radius continuity, because variance from the deepest level is greatly magnified 4/4

[Tài @LeLocTai]

I happen to be working on the exact same algorithm. Some tips I have which I've also posted on mastodon: Upscaling pass can be skipped entirely or be *much* cheaper by doing all the smoothing at the bottom. 1/n

[Kostas Anagnostou @KostasAAA]

"Fast blur with animated radius", examines various blurring techniques with an emphasis on varying blur radius and non-uniform horizontal & vertical blur. Includes WebGPU sample to showcase the techniques. https://t.co/XdIYVvr5vs

[Tài @LeLocTai]

Using 2 wide box for downsampling is cheap but can alias severely. Wider down kernel is much better but also is much more expensive. It is a bit compensated as you need less passes on average, but this is mostly for when you care about temporal stability 3/n

[Tài @LeLocTai]

@SebAaltonen This is what I did for Translucent Image. Big Yay.

[Tài @LeLocTai]

@bedrifa Is that Translucent Image by any chance? Do you want a key for Paraform? I think it fit the theme

[Tài @LeLocTai]

X doesn't handle pasted link apparently. Here's to @bgolus work: https://t.co/Y4Q5lBU51r . It's good and fast if you have high-res input and don't need smooth normal

[Tài @LeLocTai]

I spent 3 months trying to solve bitmap->sdf, and it's finally taking shape - literally

[Tài @LeLocTai]

I started this work thinking this was a simple, solved problem. You just jump flood, maybe with a little bit of subpixel correction like from our lord and savior bgolus (he's a deity back from the ancient time where people used to use internet forums to solve technical problems)