iiinova — where your digital self keeps evolving

i guess i mean specifically the first trilogy, tgaa ost goes crazy but it's not a very fair comparison

010

yeah that version is great, I like this arrangement I'm playing (sheet music boss) cause I love strides and playing across the whole piano register

020

no it's not cause of the game engine it's because it's the best song from ace attorney

0370

Acerola @acerola.gg · 14/08/2026

Been working through Godot's theme from AA3 the past couple months it's really hard and painful

4913

it's okay if the game who secretly wishes to be a movie is good but making a good movie is hard already

130

Acerola @acerola.gg · 14/08/2026

please watch my talk thank you

3894

Acerola @acerola.gg · 09/08/2026

I'm going to play fields of mistria now tho and start writing the video script thanks for looking at my silly photos I've been trying to create my own references now that I have a nice snazzy camera and also you never know what images online were ai generated now

2520

Acerola @acerola.gg · 09/08/2026

Hopefully this can give you a better appreciation for depth of field shaders that have existed over the past 15 years as they do a pretty good job approximating that circle blur for far less investment through a lot of different methods which I'll implement another time

1560

Acerola @acerola.gg · 09/08/2026

While the optimizations have been cool so far they only apply to separable filters it's great the 41x41 gaussian can be reduced from 4ms to 0.04ms on a 4090 but as you can see the 41x41 circle blur just mega mogs it aesthetically, which sadly stays 4ms as it's not separable.

21092

yes these all perform quite similarly to the real full resolution gaussian

030

not really maybe if you only did down/up once but this is 6 passes of smoothing. Only at 20x the default sample distance does artifacting become visible in max contrast

110

There's infinite variations of the dual kawase blur, the idea behind it has existed since around 2007 I implemented the arm version, but I found that downscaling to 1/16 was too much so I stop at 1/8 with a 1.5x increase in sample distance to make it identical to the other blurs

1220

The dual kawase blur is the industry standard algorithm that works the same as the kawase while also downscaling the render and upscaling back to full resolution each pass (O(n) -> O(logn)) On 4090: 0.04ms On 1660: 0.42ms (at 1080p) a 50% increase over the original kawase blur!

2843

well in this case you have 3 aesthetically equivalent blurs, one of which is significantly faster than the others, so the heuristic is obvious. In other cases, like if you wanted to depth of field or something, these blurs would be objectively bad as they do not look like a bokeh blur.

020

maybe at like, 1500x1500 an fft would be faster.

020

well that's cause bokeh functions entirely differently from these blurs. bokeh is not gaussian, a single pass box blur gives square bokeh

120

These are all nondestructive optimizations so far as no downscaling/upscaling is involved, so I'm excited to see if the dual kawase shows a measurable improvement on the 4090 considering 0.1ms is already pretty fast but it'll definitely give a big boost to the 1660

3410

After separation these 2D blurs become much faster obviously but the kawase blur reigns supreme with a 50% speedup over its effective gaussian equivalent even after separation On my 1660 though it's Gaussian: 2.0ms Double Box: 1.7ms Kawase: 1.04ms

1400

Acerola @acerola.gg · 03/08/2026

As we all know, successive box blurs approximate a gaussian Below is a 41x41 gaussian with std deviation 7.0, the double box is a 17x17 kernel, and the kawase is a traditional 5 pass Gaussian: 0.19ms Double Box: 0.14ms Kawase: 0.1ms (at 1080p on a 4090)

21151

the kawase blur is a classic for sure

060

Acerola @acerola.gg · 03/08/2026

even my 4090 kneels to a basic naive 2D box blur, a 40x40 blur kernel costs 4.3ms at 1080p yikes

4880

Acerola @acerola.gg · 02/08/2026

a nitpick I had with compositor effects was the parameter list bloat tacked on to each one causing a lot of clutter you can actually just remove them by overriding _validate_property and disabling the flag that serializes them though

1861

that is just perspective warping

050

working on something new yippee

131041

certainly not lost technology, it's actually the most basic technology when it comes to water shaders

070

Acerola @acerola.gg · 22/07/2026

not to sound pretentious but beyond the wave mesh itself this is basically just gpu gems chapter 1 which I covered 3 years ago in my video turning sine waves into water

170

Thanks so much for the kind words!

030

Acerola @acerola.gg · 21/07/2026

Omg that's me, thank you so much I'll do my best to swing by there!

130

Acerola @acerola.gg · 14/07/2026

thanks for letting me know, probably never fixing those though rest in peace acerolafx preset comparisons.............

000

Haha yeah I probably should've read more about the team background i wrote entirely just off games released.

130

I agree, but the target audience of the sponsor is not you, rather it is people who wish to expand the localities their web app supports, something google translate already implicitly does for all web devs. I do not think anyone will be using it for their video games.

100

It's machine translation technology that has existed for many years, if you think google translate is evil that's fine but the sponsor is fully within my principles.

160

i'm pretty happy it coincidentally finished on a weekend for sure

060