avatar-optimization

[somber raven]

All of that will have an impact on your Avatar. I suggest you read this article I linked: https://www.poiyomi.com/blog/2022-10-17-texture-optimization

Also, when it comes to Mipmaps, it’s not worth it to turn it off because it’s gonna make your texture look horrible. I’m serious, keep those Mipmaps enabled!

[near elk]

Thank you!

[near elk]

After reading the article, I am still confused about this part:

Let's say I have a 2k texture.
Normal Quality.

If I use Crunch Compression on it, will it appear as if I did not use Crunch Compression, when in the game?

[somber raven]

The appearance of the Texture will only be different if you use Crunch Compression or not.

However keep in mind that Texture Memory (VRAM) will always remain the same no matter what, Crunched or not.

[near elk]

If I only consider Download size of the avatar, using Crunch Compression will not lose any resolution when going in game?

[somber raven]

The only thing you will lose is a reduction of the quality in the color spectrum. If you use Gradients, they will begin to look horrible when crunched.

[near elk]

so for anything that has gradient colour I should not crunch, and instead strive to use high quality compression with lower resolution?

[somber raven]

I would say so, yeah. At most a 1024 Texture set to High Quality Compression would consume at least 1.3 MB.

At Normal Quality, it would be 0.7 MB.

However, if your 1024 Texture is using the Alpha channel (RGBA), it will always consume 1.3 MB even at Normal Quality. This is because Alpha consumes slightly more than an RGB texture not using Alpha.

[near elk]

since my goal is to compress for Quest using Toonlit(no alpha layer), the format should be any of the RGB options?

nvm, it seems by changing the format, the option to crunch compression is gone

[somber raven]

Preferably yes if you want to keep your size down. Either way, if you use High Quality Compression, the Alpha channel will already be ignored on Quest and will show as White.

[near elk]

by Gradient, would you elaborate on what it looks like? I have no experience with these

[somber raven]

Something like this as an example:

It's the blending between different colors that you gotta watch out for

The main thing to also consider is if it's really necessary when looking at the Avatar from a distance. From a distance like this (screenshot below), it's barely noticeable until you get up close to it. So it's really your decision if it's necessary or not.

Think about this when considering how people look at your Avatar in-game.

[near elk]

I think I get it.
First I'll be thinking if that texture will be viewed from a distance or up close,
Then choose the compression method.

[somber raven]

Yup! That is what you'll have to do.

[near elk]

Thanks a lot

[somber raven]

[snow knot]

I personally don't mind the slight increase in download size that not crunching your textures causes in favor of having cleaner textures. Especially on quest where you're usually less textures than what you'd normally see on pc w/ shaders like poiyomi.

Ofc that mostly ends up being personal preference. Though if you hit the 10mb limit for quest uploads you're basically forced to crunch. Then again, if you're hitting the 10mb limit you probably have other areas you can optimize other than just texture compression.

[opal garden]

Hey, i packed my Avatar quite full, but most stuff is disabled, my Animator is also shrinked down to not that many Layers, the VRAM usage is also Fine and shouldnt exeed my 4090s VRAM, any plan what could cause the Performancedraw in VRC? (about 106 FPS in an Optimized Avatar, when switching it drops to 78-86 FPS. I Speculate it could be because the way i combined the Animation layers, but ill be honest i have no plan, i could probably just split the content on the Avatar to multiple Avatars, but i dont wanna do that, i would rather like to finally find out, what exactly draws the Performance. I Also Reduced the Active Physbones a bit by Disabling the Scripts with a very minor improvment.

I wonder if someone stumbled across something that may applies to this case

i made an assumption it is maybe the whole ton of Shapekeys on the Avatar but im not sure how big the impact is

[snow knot]

I'm honestly suprised your vram usage is so low, like if I'm reading that right it's just ~50 mb of vram usage by textures.

The things actively affecting your performance here are probably the various meshes and materials you have on here. Pretty sure each of those is an additional draw call. And in the current unity version that vrchat uses blendshapes each also take a little bit of performance each, so that number looks a bit scary

[opal garden]

i tested the Usage of Disabled objects and there seem no difference in Perfomance when removing all the Objects, so it seems not to create Drawcalls

[snow knot]

Draw calls should only exist for objects that are active. So if they're disabled most of the time thats probably fine

[opal garden]

yeah, ill try out a version with the Blendshape Import disabled i guess

[bold wharf]

How many constraints do you have?

[opal garden]

about 39

but most are disabled

might draw a bit of performance overall but it shouldnt be that much

[heavy knot]

material slots 297????

[opal garden]

yeah, is that a problem?

[heavy knot]

yes, each material slot is a draw call in itself

[opal garden]

nop

not if disabled

about 266 of that is disabled

[heavy knot]

still having to offload this stuff to ram

those disabled objects have to go somewhere

[opal garden]

yeah, but the vram usage is not that big

the meshes are optimized and the textures are reused partially

and part of that is particles systems aswell

(for example) having a 64x64 texture cached from 20 Material Slots should not affect the performance and neither the Vram that much

[heavy knot]

it's still an active zip bomb essentially, you can use the profiler to see what takes the longest to render I guess

[opal garden]

zip bomb?

[heavy knot]

This probably makes system memory spike to multiple gigabytes worth of data

[opal garden]

how that?

[heavy knot]

disabled objects need to get stored somehow

system ram is where it goes

[opal garden]

empty scene with Camera and Light in Comparison:

i may remove some of the larger objects and replace them with low poly, but this is not what my questions are about

even if its in memory it shouldnt draw frames aslong as the Ram or the VRam arent full

[opal garden]

meanwhile "Gigabytes" is just wrong unless the profiler is broken

[heavy knot]

you're primarily going to need to use the profiler with something like gesturemanager active and read the frametimes and see what takes the longest to render in this type of thing. With an avatar with this many stuff just packed into it, the 202 audio sources probably don't help either even if they're disabled.

VRChat needs load in background enabled now for audio too so that's probably the biggest lead you have

[opal garden]

here with gesturemanager on

[heavy knot]

you'll need to look through CPU usage graphs

[heavy knot]

select a frame and read what took the longest to render

[opal garden]

i already looked though there but have no clue what to look for exactly

also important, this is with shapekeys removed, i havent tried that version yet in vrc, ill do that in a second

i did found out that in my current unity Project compared to an empty one, the Animator is causing more performancedraw, regardless of the states beeing just sitting on the side not connected with any transition to the Entry or any other possibly runnable State

More to that Story:

The tests where executed without the Gesturemanager

and just on an gameobject

[stuck mist]

Can someone help me with this? I don't know how to edit physbones since I'm new to unity. Also, just some general tips?

[opal garden]

oh thats quite a few, so when you click on "select", it should select Physbones in the Hierarchy , those make your Bones Wiggly, if a Bone has a Physbone Component, all its Children(Underlying Objects) count as a Physbone Transform as far as i know. You either can remove Components from Bones, that you dont need to have to wiggle, or you maybe need to take the Model to Blender and use the CATS Plugin to merge Bones together., the last option is probably the best but can be hard for a beginner when they have to navigate though Blender thier first time

[stuck mist]

I'm going to hope the first option works since I don't use blender. Am I just screwed if that doesn't work?

[opal garden]

hmm, at worst you dont have parts of your avatar not wiggle

[stuck mist]

I'm not worried if it doesn't have physbones, it already does on PC so a lot of my friends can see. The main problem is the clipping.

[opal garden]

yeah i mean without physbone no Collision so more clipping espacially on the hair

[stuck mist]

Ooooh that kinda makes sense.

[opal garden]

i was bored so i made this graph distracing myself from my insolvable unity problem

i would do one for the blenderpart aswell but thats a more complicated one, idk if theres tool for that yet in Unity, i wish there would be

...well,
i cant give you the invite link, you need the CATS discord server

the addon is fairly simple to use, you may ask someone to help you via screenshare

[stuck mist]

Is that for blender? Or unity as well.

[opal garden]

only blender

it allows you to select the bones in editmode and merge them with the Upper Bone i think in one click

i would lookout first if you can remove some physbones to see how far you can come without blender

because 200+ Transforms are a few

[stuck mist]

I could ask a friend if their is a way I could use blender, but I'm running on a laptop made in 2015 with 4GB of ram.

[opal garden]

oh hmmm

[stuck mist]

I've only been using unity this entire time but I've gotten surprisingly far.

[opal garden]

idk maybe it might somewhat work aslong as you dont render it in blender, blender is quite performant

[stuck mist]

I mean other than the massive forehead she's looking good so far.

I could always look for quest made hair.

Ok, I reduced them to 90 by sorting them out in the hierarchy I'm getting some where.

[tawdry drift]

if you can run unity you can run blender to reduce bone count lol

[stuck mist]

I've got them down to 60.

[opal garden]

just in case someone has a clue or knows more about the profiler, i would be happy to find an answer to my question : 0

it seems compared to a smaller avatar, the Process Animations Job is very time consuming

becomes way smaller when no States are in the Animator, compared to unused states...

do exit times cost more performance compared to a Condition without exit time?

[plush kestrel]

What would be an example of decent texture memory size?

So not super optimized but not straight up awful

[heavy knot]

40mb-60mb is fine

[plush kestrel]

For PC or quest

Also hmm thanks

[heavy knot]

I usually fit it in that range with most booth avatars after I put them to 2k textures instead

this is PC primarily

usually quest avatars barely have much textures to fit in the 10mb upload space anyway

never use 4k unless it's an avatar that only uses basically 1-2 main textures

if you need small details on PC like a tattoo in one place, would recommend using decals on poiyomi instead

also about decals, poiyomi will get a feature to support easier placing of decals https://vxtwitter.com/kitsuneAnCr/status/1678681713189478401?s=20

[plush kestrel]

Thank you!
Yeah I have a few models that use a single 4K texture for the entire model, and when i set the compression to none it looks perfect but the texture memory size bloats up pretty extremely. Though the moment I add any compression the texture even on the high setting it just becomes destroyed. Super blurry and the colors clump together. Its super frustrating because i have to chose between texture memory or good looks

On PC its like whatever, but I feel so bad for quest users who actually have a super limited amount of vram

I want to give them a beautiful model idk

[snow knot]

My rule of thumb for 4k is you should only use 1, maybe 2, 4k textures per avatar if you're using it for general use.

That's already like 42mb of vram just for two 4k textures

[lone tiger]

4k high quality compression and quality already destroyed? At what distance you are inspecting your model? Are you sure you are not using Crunch compression?
Personally I find DXT1/5 aka PC normal quality preset to be degrading textures the way you described, but BC7 aka high quality barely reduces quality unless you zoom in way too close.

[somber raven]

Honestly you can get away with plenty with 2048 at High Quality.

Heck, even 1024 at High Quality is damn good too. Only 1.3 MB of VRAM.

[snow knot]

Yeah, I think the only time I've used a 4k texture on an avi is when I atlased a couple materials together that uses 2k textures

I also tend to use bc7 instead of dxt5 on most of my avis since it preserves gradients better at the same vram cost. The only cost being a slightly higher download size since it doesn't compress as well iirc.

That ability to compress well is why dxt1/5 sometimes look a bit crunchy. They modify the colors of parts of the texture very slightly to compress better.

[eternal garnet]

Man, everytime i come a cross a post of some juicy performance data for stuff i get excited, just to be super disappointed.

Oh nice, raw data on how much constraints cost, i've got like 1-2 constraints. Measured data is in the 1000's to make any noticeable difference. Oof, okay. Sad.

Oh nice, blendshape costs VRAM, my avatar has a lot of blendshapes so i ca.... oh... at ~8k polys affected with 100 blendshapes its only 42mb.... so i could maybe save idk ~35-40mb if i'm really lucky. I had hoped for a bit more... i had hoped for that to be the ultimate solution to my VRAM issues... nope. Everyone would have to do it, if i do it its not gonna save me from anything.

Texture sizes and compressions, aw. 0.5mb here, 256kb there on an avatar that hardly uses 20mb to begin with.

File size ah something i.... ohwait my avatar is only 8mb big.

Animation Controllers! ... again measured in 100's and 1000's when you can only have 5 controllers on you. Layers! Nope only have like ~20.

Physbones! Of course, 100 physbon.... nope they are quite cheap with 1000 bones using 0.66ms.

Receivers/Senders those must be expensive right? 0.75ms for 1000 receivers/sender that are actively doing stuff. And i'm pissing my pants here when i have 5, they don't even matter.

[tawdry drift]

doing a lot better than most lmfao

[eternal garnet]

Trying to squeeze out every tiny 0.001ms out of my avatar when there are avatars running around with thousands and thousands of physbones, constraints and other crap.. i wish those could be optimised!

I wish we had a "strip" feature that simply strips any components past X amount.

[eternal garnet]

but i want MOAR

Also, next time i hear "omg dont use constraints, they are bad and slow" yea... no if you use 1000+ of them (how the hell do you even manage to get 1000 constraints, how the heck do you even set that up, how the heck do you even keep track, it must be hell to navigate)

there will be heads rolling

[tawdry drift]

theres a centaur avatar that came out recently :^)

uses 100s of constraints iirc vaguely

very cool avatar but dam

[eternal garnet]

but thats like idk 2 per leg?

how the hell does it use 100

[tawdry drift]

keyword vague, i haven't digged into the package yet bc its 70usd but like

from the skeleton look on the gumroad page theres 3 different humanoid skeletons its operating from

took like 3 years for the creator to make, pretty cool

[eternal garnet]

different skeletons... how the hell do you even do that?

or do you mean simply 3 armatures

[tawdry drift]

some smaller stuff like reducing blendshapes aren't gonna matter once we go to the next unity version though

[eternal garnet]

and why 3 separate ones...

[tawdry drift]

yeah 3 armatures sorry

im a little smooth brained atm lol

[eternal garnet]

i have so many questions...

[tawdry drift]

hoping the bot doesn't yell at meeee

https://voxian.gumroad.com/l/AvaTAUR

[eternal garnet]

1,350 Animations | 1,250 Animation States | 5,400 Animation Transitions | 235 Blendshapes | 28 Hair and Body Features | 90 Menu Options

wth

1250 animation states

5400 transitions

https://tenor.com/view/going-crazy-dizzy-exhausted-cant-take-it-enough-gif-15915942

[tawdry drift]

it does hit the fps but the creator is working on optimizing

[eternal garnet]

i think my heart just skipped a beat or two

[tawdry drift]

ive tested the public version and its super cool lol

[eternal garnet]

calculating the velocity of the hind legs via a set of 9 contacts and a few constraints

what...

and then theres me be like OMG i'm so scraping it with these numbers

all these basic anime avatars i see with 200mb VRAM usage

me with highly detailed body + normal map + ao + specular + emissive + 2 entire outfits

how do you even achieve 200mb on an avatar with a single color skin

[tawdry drift]

love that design but probably masks, matcaps, shit like that

and normals too

ive only ever had 1 instance on a robotic avatar where increasing the normal map size improved the look, but otherwise you could get away with 512x512 normals

a lot of those very poor gumroad avatars are all using 2-4k for everything

anything that isn't the diffuse texture can be 1k or lower

[eternal garnet]

i'm using 2k for the main texture and the normals

and those are already stretching it

i had to manually scale them down in GIMP (to get the best resize algorithm) and save them as TGA (since it improves quality very slightly over PNG in Unity)

i also baked out some AO from the normal maps to add some more shading and color variations to the fur pattern, so going below 2K on diffuse is absolutely not an option

but what pisses me off the most is the BC7 compression

its such a wonderful format... except i can't force it to alpha off, so in the personal mirror i end up with slightly transparent freckles on my avatar all over when a bright glowy background shines from behind the mirror

the only way to get rid of that is choose an alpha-less compression (DXT1 for instance) but then the quality drops so hard that i might as well be setting it to 1k and enabling crunch compression

not to mention on my wicker at least i cannot use an alpha-less compression as i'm using the alpha channel for my emission map (i'm doubling my diffuse as emission map by reading the alpha as emissive intensity)

and thats already some 1337 haXX0r as i had to edit Mochi's Uber shader to even support that

What i essentially did was set my diffuse texture up to alpha the entire body except the emissive parts, non emissive parts had to be at least ~4-5% opaque otherwise the compression would absolutely destroy the colors and introduce insane color banding, so i had to add a cutoff value in the shader on top of making it read out the alpha value, anything below 5% is simply ignored and seen as non-emissive

just to save a single extra texture lol

but that reminds me, since BC7 does a piss poor job at deciding which chunks of the texture should be compressed with alpha and which don't (even in a fully solid texture it does that), i wonder if reexporting the texture in a format that doesn't even have alpha helps (or making sure that i remove the texture's alpha channel in GIMP before exporting)

i hate this

[rain matrix]

Because your avatar isn't built around PBR

I agree that 200mb is way too much

But you also can't draw a comparison between a toon shaded model and a full on PBR workflow model

[ivory sluice]

Pbr doesn't depend on texture resolution tho, it depends on the shader

[buoyant holly]

heck you could technically have a PBR model with no textures just setting values in the Shader

[rain matrix]

That's true but more often than not, if you want it to look good and specific to your usecase, you need a lot of maps

That's why I said pbr WORKFLOW

Aka baking all the maps needed for PBR

And depending on what you do, those can stack up quickly

[buoyant holly]

Channel packing can help reduce the amount of textures needed

like if it doesn't need metal you only need two textures your color map with the smoothness embedded into the alpha Channel and a normal map

[rain matrix]

Yes thankfully

But really my point is that it's A LOT harder to optimize a PBR workflow

It's a really deep topic

[buoyant holly]

if you're using substance painter it's not that hard to set up the export settings to be efficient Channel packing wise

[rain matrix]

And a lot of people don't possess knowledge that deep to optimize it to its fullest

And say fuck it

And this is also the point where I will start shifting blame onto the performance rank system for encouraging unoptimized avatars

[buoyant holly]

detail Maps can be handy for reducing the amount of texture resolution requires https://twitter.com/BlizzyFoxo/status/1655354885926264834?s=20

Blizzy (@BlizzyFoxo)

@poiyomi shaders have the ability to use a color mask for detail normals. I really wish more people would use this. You can use absolutely tiny maps and still have the detail you want. 10 times less VRAM usage for the normal map on this sweater INCLUDING the mask!

Likes

107

Twitter · May 7, 2023, 11:32 PM

[buoyant holly]

out of curiosity which bit of the Avatar ranking system do you find leads to counterintuitive results

[rain matrix]

A lot of it but it's primarily poly limit

You reach 70k so quickly

And once you hit very poor why bother optimizing anything else? Might as well go all out if you're gonna be Very Poor anyways

So people just straight up don't care about optimization cause they're gonna be VP anyways due to poly limit

[buoyant holly]

wonder if some of that triangle count could be optimized if people use normal Maps more

[rain matrix]

Yes but also

Why bother

That's my exact point

[buoyant holly]

also probably doesn't help that folks try to cram multiple outfits in one Avatar

[rain matrix]

There's really two main issues in my eyes and they kinda go hand in hand. VRChat simply does not provide even basic tools or guides on optimization, yet expects people to optimize

You can make tools in Unity to help people reduce poly/material/mesh count

To help them optimize animators

Particle systems

People shouldn't need the equivalent of a 4 year 3D modeling degree

[buoyant holly]

yeah they could definitely update the section about Avatar optimization with some more tool recommendations

[rain matrix]

They should straight up provide these tools

In the SDK

People aren't gonna go out of their way to go to the docs, find that specific section, then go get that tool

[buoyant holly]

I guess a question would be even if those tools were in the SDK what would incentivize those people to use them?

[rain matrix]

Not being VP

By pressing 2 buttons

If you hit the 70k poly limit and there would be a button next to the performance stat that would open a window where you can remove meshes and decimate them individually and save that as a new version of your avatar, fully set up, a lot more people would start to care

The less resistance the better

The majority of people simply buy avatars to upload them, not to mess with Unity or Blender

[buoyant holly]

I guess that would depend on the original mesh as far as how much of a messy Auto decimate feature would do

[rain matrix]

The point is that it is an option

Fuck it, give more advanced users a window for simple barebone retopo to adjust the meshes after decimation

The less external tools/programs are needed the better

There's A LOT of room for what VRChat could do to actually help people optimize

[heavy knot]

Polygons is the least of your worries with optimization unless the mesh is caked in shapekeys, people are mistaken that polygons are the worst thing to exist because it’s the first metric they’re met with on the avatar stats screen. Material Slots and Animator Layers are the main CPU chewers in the mix, most modern GPUs especially ones that run vr well can handle quite a lot of polygons tossed at them.

Even then if you’re running multiple outfits and they’re actually disabled, you should be fine as they’re offloaded to DRAM in the meantime.

[marble rain]

It does not offload into memory. It stays in vram til it's no longer needed.

[marble rain]

Not exactly true. If you have 40 people with 100k poly each. Your loading in 4 million that. Which is quite alot.

[heavy knot]

I’ve seen a 1 million polygon with one material slot potato run better than an avatar with 20 material slots

[marble rain]

And it's easily testable to out in 40x with 70k and then 100k for example. There is going to be quite Abit of a difference

[heavy knot]

You’re CPU limited in VRChat most of the time

Not GPU

[marble rain]

Not really

[heavy knot]

If you’re getting high GPU usage, you’re either running a x3d cpu or not having an up to spec gpu from the past 8 years

Or the world is battered in post processing

[marble rain]

With even just 20 people I am looking at 100% GPU use

Nope

[heavy knot]

What gpu/cpu combo are you running

[marble rain]

Rtx 2070 and it 10600k which is plenty powerful.

Nothing but GPU struggles

[heavy knot]

It depends on what the avatars are doing with shaders primarily then

[marble rain]

Not really no.

[heavy knot]

People use grabpasses for things like glasses when they shouldn’t

[marble rain]

Eh most dont

They just use transparency

[heavy knot]

Is your metrics ran in vr or desktop. If it’s in vr then what headset

Resolution and msaa can be a real hitting factor

[marble rain]

Beside that. The problem in vrc is the amount of vram needed cause vrchat still does not use single pass I believe

[heavy knot]

Yeah they don’t, making double draw calls still an issue

They got unity to make a build without it on unity 2019

[marble rain]

They wont

[heavy knot]

Seems like they’re forced for the 2021 upgrade

It’s probably 2022 now

[marble rain]

They are going 2022 which is better

Just don't shoot above that 70k for avatars. It's easy to make alot of things fit within that

Obviously keep mats down to max 4

Physbones to 64

Etc

[heavy knot]

Going a bit higher is completely fine in term of polygons, just not 300k+ then it just gets horrid

I can see that

I Usually get the space of 70k-140k anyway

[marble rain]

That's sadly where things struggle real badly I have seen it to many times

[heavy knot]

The performance ranking system is just a meh attempt at most

[marble rain]

As soon as I load in. A 140k it just drains my frames lol

[eternal garnet]

All my avatars are made for PBR, they just don't use metallics maps (because quite franky a simply using global shine just super toned down and very very soft is enough on a model that is supposed to be "fully furred").

I was referring to how do simple, single color anime avatars, that have 0 detail manage to do 200mb VRAM usage when my "full PBR" avatar that has a diffuse, normal, (sometimes metallics for my armor) and emissive can stay all the way down to ~25-40mb, with some 2K textures mind you.

the only way i could imagine ever achieving that is using a simple 1 color (that could be as big as 1 pixel) that is 4K or even 8K and disabling texture mipmaps for practically everything.

heck i mean enabling texture mipmapping and streaming alone massively reduces your VRAM usage even with 4K on everything, since 4K already fades out to 2K at an astonishingly small distance (we're talking barely a meter away), so you'd kinda have to purposely disable mipmapping to get there... unless you have like 20 outfits on that avatar...

and even then, as far as i could tell, disabling body parts/objects will reduce your VRAM usage with mipmapping enabled as it seems like unity falls back to the lowest texture (if not outright dropping it out of VRAM into your RAM) when a material/texture that has mipmapping enabled is not visible.

[rain matrix]

Mipmap levels don't influence your VRAM

All mip levels get loaded into VRAM at once and stay there

[eternal garnet]

doubt

[rain matrix]

Doesn't matter how far away someone is, their full res mip stays in VRAM

[eternal garnet]

the point of mipmapping + streaming is that you only load the texture level necessary, that saves VRAM

everything else is loaded into RAM

[rain matrix]

Streaming is disabled ingame that's why

[eternal garnet]

is it tho?

[rain matrix]

The SDK forces you to turn on streaming in case VRChat ever decides to utilize it

[eternal garnet]

According to world tests it works

[rain matrix]

But the actual game itself does not make use of streaming no

All mip levels are loaded into VRAM at all times

[eternal garnet]

so you are saying that the VRAM stat is completely and utterly broken and wrong

[rain matrix]

I mean it doesn't count mesh VRAM anyways

[eternal garnet]

thats not what i'm saying

[rain matrix]

So it's already lower than it actually is

[eternal garnet]

you are essentially saying, enabling mipmaps is bad

[rain matrix]

I'm unsure if VRChat calculates that into their stats or not

[eternal garnet]

it does.

[rain matrix]

No

[eternal garnet]

thats why im saying that

enabling mipmapping and streaming does lower your VRAM stat

quite massively actually

its how i lowered from 120mb down to 40

[rain matrix]

Mipmaps still speed up the rendering process and increase visual fidelity

Enabling mipmaps saves a lot of frametime

[marble rain]

Mip map is useful to reduce load the further away you are

[rain matrix]

Yeah

[eternal garnet]

at the cost of more VRAM usage

[marble rain]

A little bit

[rain matrix]

Also textures that are further away look like trash without mipmaps

[marble rain]

Yee

Even close by

[eternal garnet]

highly detailed and repeating ones

yes

[marble rain]

It's only 5.4 mb for a 4k Tex btw

To use mip map

[rain matrix]

But yeah it still saves you a metric ton of processing power even tho it takes up more VRAM

[eternal garnet]

disk.

[rain matrix]

Keep in mind that VRAM usage doesn't affect performance until you start hitting the ceiling

[marble rain]

Huh

[eternal garnet]

1K is 4MB in memory

2K is 16, 4K is 64

[marble rain]

Eh I am saying mip map mate lol

[eternal garnet]

yes mipmap but you are referring to disk space

[marble rain]

A 4k takes 16mb

[eternal garnet]

or file size

[marble rain]

And no

That's vram

[rain matrix]

Mipmaps without streaming is around 30 % more VRAM than without

Regardless of full res

[marble rain]

21.4 mb for a 4k mip map enabled tex

[eternal garnet]

no.

[marble rain]

Yes lol

[rain matrix]

[marble rain]

I hope you know that everyone uses compression rena

[eternal garnet]

1024x1024 is about 4MB in VRAM, due to colors and channels. 2048x2048 which is 4 times as big in pixel amount is roughly 16MB then, 4K which is 4 times 2048x2048 is 64MB

compression doesn't matter in VRAM

[marble rain]

Yes it does lool

[eternal garnet]

textures get uncompressed in VRAM

[marble rain]

Nope

Compressed textures do not uncompress in vram.

No engine does that

[eternal garnet]

not the engine

the engine doesnt matter

[rain matrix]

I think youre confusing crunch compression with compression lol

[eternal garnet]

its how rendering and GPU's work

crunch compression is solely disk space

[rain matrix]

Crunch compression gets uncrunched on the GPU

Image compression stays the same in VRAM

[marble rain]

No one talks about crunch compression lol

[rain matrix]

You can, right now, take a 8k texture and turn off compression

And it will take up 240mb of VRAM

Instead of whatever it is when compressed

[marble rain]

Yee

The compression is pretty neat. Saves so much vram

[eternal garnet]

It really doesn't.

[rain matrix]

What 💀

[marble rain]

Lool

Renamon please just Google it

[eternal garnet]

A texture in memory is always width x height x bytes

[marble rain]

No

[eternal garnet]

what i just did.

[rain matrix]

You saying going from 240mb to 20mb isn't a lot?

You're confusing crunch compression with image compression

Two entirely separate things

[marble rain]

A 8k compressed red is 64 mb btw

[eternal garnet]

im not talking about crunch compression, i dont even use it.

[rain matrix]

You can then further reduce the image disk size from 20mb to whatever using crunch

[eternal garnet]

because you are stripping the green/blue channel

[marble rain]

Your confusing it with crunch

[eternal garnet]

im not.

[rain matrix]

Ah damn

[marble rain]

You are lol

[eternal garnet]

also a red only compression is a special case

[marble rain]

What are you on about rena?

[eternal garnet]

if you strip channels out you end up with less bytes per pixel

[marble rain]

They also don't do that

[eternal garnet]

i'm saying, a texture under normal circumstances is uncompressed into memory, that is pixel width x pixel height x bytes (usually 4 unless you strip channels). 1024x1024x4=4194304 bytes = 4MB.

[marble rain]

Your litterly saying the most widely used way to reduce vram In games is not working

That does not matter.

[winged palm]

Hi all, I'm considering getting a portable monitor for my laptop. Can anyone share their experiences or recommendations?

[marble rain]

The size is not being reduced. It's the bpp that is reduced

When Unity loads a texture with a compression format that the device supports, the GPU can use the data without any need for conversion.

Quite litterly says right there on unity's on site. On how it does it

https://docs.unity3d.com/2020.1/Documentation/Manual/texture-compression-formats.html#applying-texture-compression

And it always depends on the engine

[eternal garnet]

yea i've seen those before

guess how i got my VRAM usage down so much.

[marble rain]

By compression like everyone does

[rain matrix]

Hope this visualizes it

[marble rain]

Thanks red loo

[eternal garnet]

well then

still tho, pretty sure streaming mipmaps work in VRChat

[marble rain]

Not at home so I can't show it atm

It does not

Streaming requires special setup

More server power btw

[eternal garnet]

then VRChat's avatar statistics falsely take it into account

[marble rain]

Nope

[rain matrix]

the mb it shows you at the bottom of the image in the inspector is the VRAM usage if crunch is disabled. Switching compression from low quality to none increased VRAM from 85 to 341

[marble rain]

The one vrc shows is texture only.

Not the complete avatar

[eternal garnet]

ugh

why does it matter?

im talking about textures

streaming mipmaps is a texture feature

we are talking about textures

so why does it matter if VRChat includes the avatar VRAM usage?

actually... it does

NOT including the avatar makes it cleaner

[marble rain]

Eh it matters a crap to. Lol

[eternal garnet]

NOT for this discussion

and NOT for my point

[marble rain]

Everyone runs out of vram.

[eternal garnet]

yes

but my point being that streaming mipmap works because it reduces the VRAM usage displayed in VRChat

[marble rain]

No

[eternal garnet]

unless VRChat is falsely taking it into account when it shouldnt

[marble rain]

It does not work

Never has in vrc

And most likely wont

[rain matrix]

They are currently not using streaming no

[marble rain]

For a while

[rain matrix]

They force you to enable it in case they ever find a way to use it

[marble rain]

Yee

It's Abit hard to make it work

[rain matrix]

Because enabling streaming at runtime for textures uploaded without that setting enabled is pretty much impossible

[marble rain]

Since you need to code it in for everything

[eternal garnet]

why does it matter? just dont use streaming for textures that dont have it enabled

[sick relic]

vrchat team even stated the people will abuse avatar texture streaming priority to make everything else looks bad.

[rain matrix]

hm

[marble rain]

Eh.

[rain matrix]

Real af

[eternal garnet]

thats why you can mix and match them in unity after all, it will use it when enabled and wont if its not

[rain matrix]

No

VRChat does not use them if theyre enabled 😭

[eternal garnet]

thats not what i said.

[rain matrix]

VRChat, ingame, at runtime, does not support streaming

Turning on streaming does literally nothing

[eternal garnet]

i said: UNITY allows you to mix and match, using it when enabled for that specific texture and not using it if its not enabled. This is an engine level thing and should not factor into their reason for not using it.

[rain matrix]

Yeah but whats your point

They cant use it due to technical limitations, so they dont

[eternal garnet]

I'll test it in VRChat.

[sick relic]

Here's the reason why they don't: people will abuse avatar texture streaming priority to make everything else looks bad.

[rain matrix]

You do not need to test it we know streaming is disabled ingame

[eternal garnet]

If it does, i'll keep using it, if it doesn't i'll disable mipmapping alltogether on all textures (pretty much all of them) that don't need it.

[rain matrix]

damn

No dont disable mipmaps

for the love of god

[eternal garnet]

Simple solution? Don't allow priority? ignore priority?

Why would i keep mipmapping around for my small textures

[rain matrix]

unless you want your gpu to be in pain and your texture to look like trash at further distances, do not disable mipmaps

[eternal garnet]

i'm just artificially infalting both my avatar size and its VRAM usage

[rain matrix]

mipmaps speed up the rendering process a lot

they still help performance A LOT

dont turn it off unless you have an actual really good reason to

and those reasons are like... one

if you have small text you want readable over longer distances you turn off mipmaps, in literally every other case, keep them on

[sick relic]

Mipmap and streaming mipmap are different things though. Mipmap is like presampled texture for cheaper texture sampling in GPU. Streaming mipmap is a technique to dynamically load mipmap level depends on available resource.

[rain matrix]

Yeah exactly

Please keep mipmaps on

Your GPU will thank you

And everyone elses GPU

[eternal garnet]

I don't see a point in keeping mipmaps around for a specular/metallic map that is hardly 256/512 anyway if its just a low resolution shine effect that is barely noticeable and acts more of a general "this plate is that shiny and that plate is this shiny" , it really doesn't matter on distance and won't look any different also a 256x texture isn't hard to keep around.

Look my GPU is already crying, i'm on a GTX 1060.

6GB is instantly 50% full just starting steam VR and oculus link, VRChat practically instantly fills it to 100% the moment i just login into any world

[rain matrix]

Because if you have a 4k texture your GPU has to render the entire texture at all times, rather than rendering only a lower res texture at further distance. The highest res version of a texture is very rarely rendered with mipmaps, but if you turn off mipmaps it will be rendered at full res at all times

[eternal garnet]

and those avatars with 300mb VRAM usage aren't exactly helping

[rain matrix]

if your maps are only 256mb then mipmaps will not make a huge difference

[eternal garnet]

Funny that, because you are in your own avatar, meaning you are as close to yourself as you possibly could, you will always see the maximum resolution anyway.

[rain matrix]

Wrong

For highest mip to render you have to be like

[eternal garnet]

inside yourself?

[rain matrix]

Mega super giga absolutely horrendously close

The second mip kicks in almost instantly

[eternal garnet]

i kinda doubt

the world tests someone made with blocks changing numbers on distance showed very clearly that thats not the case

[rain matrix]

Also turning off mipmaps because you think you will render yourself at full res at all times is hella egoistical

[eternal garnet]

its around 1m~

[sick relic]

Mip level also effected by viewing angle. If you view a polygon from near parallel angle, the mip level will be much lower.

[eternal garnet]

hella egoistical, haha, after all my squeezing and extreme optimization to tickle out the last few kilobytes out of my avatar and you are seriously holding me up for that?

[rain matrix]

Ah man do I love having to process much more than necessary because someone wants to save a tiny bit of VRAM, causing a bigger performance impact for everyone, including themselves, than mipmaps would

Amazing

[ivory sluice]

Mip maps are important and it's true that vrchat doesn't have streaming mip map support right now

But mips don't help on gpu performance, no matter if your texture is 1x1 or 8kx8k, the process will be the same, for each covered pixel on the screen, the gpu acesses the texel on memory and samples it, one or 4 times depending on if it has filtering, it would be amazing if yiu could tell the gpu to sto acessing memory for a certain texel until it changes to another one but that's not the case

So textyre sampling will have the same cost no matter the resolution

Also don't use 4k textures unless it's an everything-in-one atlas

[eternal garnet]

boom

also once a texture is allocated in memory, it is allocated, allocating and deallocating is the slowest process.

[sick relic]

Optimization is a tradeoff between quality and performance. Currently you're at the point where you have to tradeoff quality and performance from mipmapping for just a few KB of vram saving.

[eternal garnet]

so unless you were to deallocate and reallocate textures constantly (texture thrashing, level changes or other on-demand new texture loading/deloading) you have nothing to fear

look i've already gone through all my textures, each and every single one, tried all resizing algorythms and determined the best for each type and resolution to use, saves them in a format that doesn't further dump down the quality on unity import and set a compression level that saves as much as possible while keeping the quality as close to the uncompressed version as possible, i have combined textures wherever i could, even modified shaders to support my hacky implementation of using the alpha in my texture for emission so my diffuse can double down as emissive map just to save on an extra texture, give me some slack.

[rain matrix]

Mipmapping increases cache efficiency as full-size textures will not be needed as often and the lower resolution mipmaps will fit more easily in the texture cache

Texture data doesn’t have to be fetched as often, reducing bandwidth usage

[eternal garnet]

According to you the full size texture will never be used.

[rain matrix]

I never said never

I said you have to be hella close

[eternal garnet]

I doubt, both Unity and in-VRChat tests both world and my avatar has shown that mipmaps start fading after inbetween 1-2m away from you. I could see that because for a while my texture had no UV edge extensions, so scaling down the texture (or mipmapping even just to the first mipmap level) would result in seams everywhere, very clearly visible, both in Unity and VRChat

i'm A very confident that having your camera inside your head is less than 1m away from yourself and B you get close to people more often than you think

[rain matrix]

Deep blue is full res

Light blue is second mip level

[eternal garnet]

yea

that isn't almost instantly

[rain matrix]

That not even 1 unit of distance 💀

Thats less than a meter

(mip distance levels depend on texture res and object scale bla bla bla)

[eternal garnet]

#avatar-optimization message

i said around ~1m

[rain matrix]

Yeah thats really close

[eternal garnet]

ehhhh

[rain matrix]

Unless you have people sniffing you constantly

[eternal garnet]

close enough for yourself.

[rain matrix]

Yes but also

Other people exist

[eternal garnet]

i do have people being very close a lot, yes

[rain matrix]

Please for the love of all that is good do not disable mipmaps

[eternal garnet]

point being you made it sound like the full res texture is pointless

which implies i could have dumped the texture one level down and forced full res and get the same result

which would be sad considering that a drop by just a single level means a massive quality loss

[sick relic]

Also don't forget that mip level is effected by viewing angle

[eternal garnet]

right i know how mipmapping works and what its used for

[sick relic]

And you still insisted to disable mipmapping for just a few vram KB saving?

[eternal garnet]

i didn't insist

i haven't done anything yet.

as with anything i will be looking very closely into this and how it affects the performance

[rain matrix]

mipmaps still save performance due to caching

[eternal garnet]

wasn't it just previously said it doesn't help performance, not to mention it actually costs performance in theory since texels need to be sampled, thus requiring some work

if i remember correct unity even states it trades some work cycles

although that may have been for streaming mipmaps

[rain matrix]

saying disabling mipmaps saves performance because of less VRAM is like saying your cpu die cache is the same as RAM

mipmaps still increase performance due to caching

[sick relic]

Streaming mipmaps saves vram by loading only necessary mipmap level into vram. But vrchat doesn't utilize streaming mipmaps yet. Therefore the texture gets allocated into vram fullres plus mipmaps if mipmap generation is enabled.

[rain matrix]

Sure, all mipmaps will get loaded into VRAM at once, but only the currently used mip level will be cached. If you're rendering someone further away, only their low res mip will get cached, freeing the cache up by a lot. Also, since the GPU needs to fetch a much smaller texture from the cache, bandwidth usage will be much lower and fetching faster

[eternal garnet]

if they are all cached and are all loaded into VRAM at once how exactly does it save performance using a lower one? You only ever swap out the currently rendered texture for another one (unless you are blending between two), so effectively you are only ever using (and reusing) a single texture which doesn't cost anything really, the cost comes from swapping between lower and higher resolution ones

[rain matrix]

If everyone would turn off mipmaps, your cache would fill up insanely quickly and your bandwidth will suffer

[heavy knot]

If you're needing to dig through couch cushions for specs of VRAM, I'd try resolving the issue first at hand like using Virtual Desktop or OculusKiller if you can't do wireless.

[eternal garnet]

Not an option.

[rain matrix]

They're not all cached, that's the point

[eternal garnet]

Oculus killer gets rid of the only good keyboard and desktop that Oclus offers. SteamVR's keyboard and desktop is absolute garbage.

[rain matrix]

If an avatar uses a 4k texture but is super far away, your GPU will ONLY cache the 16 x 16 texture for example

And when rendering that avatar it only needs to fetch the 16 x 16 texture from cache

Rather than fetching the 4k texture on every call

[eternal garnet]

why would you cache that texture, you have it actively allocated already, caching is meant for things that would need to be recalculated otherwise, a texture that lies in VRAM doesn't need to be recalculated, its already loaded and is used when necessary

caching would be useful for instance if you had to calculate vertex transforms every frame (due to a blendshape) on an otherwise static object that doesn't have bones or changes its shape otherwise, so you can reuse the cached data

[marble rain]

Cache is already allocated. Its the most frequently used data that gets stored before it moves to onboard

[eternal garnet]

but on a texture.... that is actively loaded into VRAM...

but why store it... its already in VRAM...

[marble rain]

Because the cache is 1000x faster

[eternal garnet]

its already at its final destination

[rain matrix]

cache is much much faster

its why your cpu has cache too

[sick relic]

VRAM is located outside of gpu core die while gpu cache is located inside of it.

[eternal garnet]

i think we are talking about different "caching"

[marble rain]

No

Cache is cache

[rain matrix]

Cache is an actual physical part of your GPU and CPU

[marble rain]

So is vram through lols

[eternal garnet]

well then i dont understand the point of the gpu cache i suppose, since the texture is already in VRAM thus its final destination, i fail to see how a cache would speed this up considering that the GPU only has to use the texture now

[rain matrix]

Because cache is much faster

Than VRAM

[eternal garnet]

oh

[marble rain]

Cache is where you store the most important and used data.

[eternal garnet]

i see.

[rain matrix]

The faster your processor can access data, the faster it can process them, the faster it is done with a task, the more tasks it can perform per second

[sick relic]

Like L1,2,3 caches for CPU

[eternal garnet]

yea

makes sense now

the active cache... i was thinking of a traditional cache

[rain matrix]

When someones avatar is on your screen it loads its textures into cache because it will (hopefully) need to access that texture to render 60+ times a second

Rather than retreiving it from the VRAM every time, which is located much further away from the processor

Which would take longer

[eternal garnet]

i was thinking why would you load something into cache, if its in VRAM, you are just loading it into cache, to load it into VRAM and then use it... if the cache has a direct faster access then of course the cache makes sense

[rain matrix]

Once all your cache is full your GPU will begin pulling from VRAM directly, or swapping data in and out of the cache a lot more often, which will lower performance

[eternal garnet]

how much is up for debate, i've frequently seen 10.000+ textures per scene being used

[rain matrix]

With mipmaps, only the currently relevant mip is loaded into cache. Meaning if someone is super far away youre not caching their full res texture, but a lower res mip level

[eternal garnet]

yea now see that makes more sense

[rain matrix]

This is also extremely important because the further away you are from stuff the more stuff you can see at once, so theres more textures needed

[eternal garnet]

the more you can see at a faster rate

please be very specific

[rain matrix]

So you dont want eight 4k textures loaded into cache if visually you can only see 16 pixels of it

Your cache would much rather prefer eight 16 x 16 textures

[eternal garnet]

yea no its okay

i see where the confusion comes from

[rain matrix]

Yep yep

[eternal garnet]

the typical vagueness that i struggle so much with

[rain matrix]

And the processor also has to fetch every texture that it needs to render on every render pass aka every frame, and pulling a 16 x 16 texture is much faster than pulling a 4k texture

[eternal garnet]

due to bandwidth

[rain matrix]

Yeppers

[eternal garnet]

again

be specific

those small details are what get me frequently confused

[rain matrix]

In conclusion, mipmaps do fulfill a very important role and despite using 33 % more VRAM per texture, increase overall performance due to caching optimization

VRAM usage doesnt cause performance issues until your VRAM is actually full, at which point the GPU will have to pull textures directly from RAM

[eternal garnet]

i've been told that i struggle understanding conversations due to a common "vagueness" being present in such conversations. You have to be very very very specific and precise otherwise ... this happens

[rain matrix]

Wether your GPU uses 1GB or 4GB of VRAM, if it doesnt have to render aka access any of it, performance will be the same

[eternal garnet]

my VRAM is always full anyway

[rain matrix]

Yeah I get that lmao

Thats generally the case in VRChat unfortunately

[snow knot]

My problem is that I alway end up unintentionally vague due to expecting the people I'm talking to to already have an understanding of what I'm talking about

[eternal garnet]

That's why i'm saying it might be more benificial to disable mipmaps on certain textures, specifically small ones, again you would be trading one for the other, it's a very very small line to walk but again i've already been squeezing harder than its worth.

[eternal garnet]

Well thats just normal, most people do that otherwise you end up handholding someone like they are an idiot. But that's not what i'm talking about

My american friends have been hinting at the fact that it has to do with me being german thus coming from a different way to articulate yourself, mind you german is a much more precise language than english and they also mentioned that from their own experience americans often tend to talk vague in general rather than going down into every little nitty gritty detail as it would overcomplicate what you are saying (and possibly lose their interest or track)

You could say it gets "lost in translation" is what its often referred to

Ohwell. Anyway. Doesn't fix my problem.

I still cannot tell BC7 to not use the alpha channel at all, preventing the transparency artifacts on an otherwise opaque texture, not without changing the compression format to something that doesn't have an alpha channel (that will either massively inflate VRAM usage or lower quality to unacceptable levels)

DX1 is absolute unusable garbage. RGB16/24 is way too big.

Anything else is not usable for a diffuse texture that needs RGB

why is there no BC8 that trades the alpha bits for more R G or B

and why is there no BC that simply allows us to manually set how many bits per channel

[sick relic]

Because compression algorithm has to be very fast and work on GPU instruction set.

[eternal garnet]

I hate how android has so many different samey compression formats with different chunk sizes

so? just make one compression format that uses the best compression algorythm but allows you to customize RGB bits and A inclusion

chunk size

just make the ultimate compression format, drop all the rest

[sick relic]

Variable bit length compression would require more complex algorithm and is slower.

[eternal garnet]

then split it

but get rid of all the other million ones.

[snow knot]

Weird. I don't think I've come across any artifacts like that when using bc7 on textures that don't have alphas on any of my avis.

[eternal garnet]

#avatar-optimization message

the avatar becomes see-through in masked mirrors for instance

anything that masks out the environment will show this when bright (emissive) surfaces are in the background

[snow knot]

Which shader are you using?

[eternal garnet]

Mochi's Uber shader

[marble rain]

Masked mirrors shouldn't do that unless the shader is well build wierdly

[eternal garnet]

i've already tried looking into it and seeing if i can somehow force it to ignore alpha but the shader is ... more complicated than i thought

[snow knot]

ah, I use poiyomi. There's probably something funky going on w/ the shader maybe

[marble rain]

Use poiyomi for sure

It's alot better

[eternal garnet]

I hate poiyomi

i originally came from Poi

[snow knot]

Debatable

[marble rain]

Why?

[eternal garnet]

its lighting breaks horribly in many worlds

[marble rain]

No lol

[eternal garnet]

namely common gameworlds like Among Us where there is no proper lighting

[marble rain]

It won't break either

I use it and never breaks

[eternal garnet]

i have not yet managed to make poiyomi look decent in Among Us

[marble rain]

You just need to set it up

[eternal garnet]

thats what i did for over a year of fiddling with options

i just cannot get some decent shading without using some hacky environment rim (that's going to show up on properly lit worlds too)

[marble rain]

Use realistic settings and then override minimum light to ensure you are brighter then the world enforces

[eternal garnet]

its not the brightness, its the shading

due to the lack of sun

[marble rain]

Eh the shading works fine to

[eternal garnet]

also i always used realistic

[marble rain]

Your prob using it wrongly then

[eternal garnet]

Again i could never get it right after over a year of fiddling with every value i could think of

it either looked "decent" in Among us or everything else

[marble rain]

Just don't go to those worlds

[eternal garnet]

but i like killing people

[marble rain]

Since their shading is bad

[eternal garnet]

simply avoiding game maps made by Jar is not exactly an option

[marble rain]

It is

All other maps he has works fine to

[eternal garnet]

well yea but i dont like those maps

i want to play Among Us

not Freeze Tag

[heavy knot]

you're just going to have to animate certain shader parameters and control them with a toggle when you're in a horribly lit world then

[snow knot]

Anyways. You use a mostly pbr setup for your avi right. Could probably check w/ the default unity shader to see if that issue you're having is specifically from mochie.

[eternal garnet]

just dont go to Among Us if you want to play Among Us isn't exactly an option

good idea but as far as i'm aware unity standard doesn't even support alpha in diffuse like at all

i mean it shouldnt

[marble rain]

Just fiddle around with poiyomi more. You will find it

[snow knot]

Shouldn't mochies also not then? Unless it's always set to non-opaque

[sick relic]

You have to set render preset to transparent or fade in Standard shader.

[snow knot]

Never really used mochies so idk how it's normally set up

[eternal garnet]

Yea, which is something you can do in Mochi too

[eternal garnet]

If set to Opaque it shouldn't yes.

[sick relic]

Isn't the point to test that there's something wrong with transparency in mochie shader?

[eternal garnet]

It is already set to Opaque

that's the thing.

[snow knot]

And you aren't even using the alpha on the image, and the alpha channel is pure white/black?

[sick relic]

If it's opaque then it isn't going to use alpha channel.

[eternal garnet]

Yes, i tried specifically stripping the alpha channel out of the texture too

Unity however will regenerate one back in

it COULD be however a mipmap issue

nope

[eternal garnet]

its not.

Yup. Didn't help.

already tried all of that

[heavy knot]

setting alpha source to none removes the alpha channel in unity

[sick relic]

How did you see unity generate it back?

[marble rain]

Yea. It does

[eternal garnet]

again i think the issue comes from using BC7

since BC7 gets to decide itself whether it uses A for a given block

this might actually override your alpha source setting

[marble rain]

Not if there is no alpha

[eternal garnet]

well clearly not!

[sick relic]

Are you sure that it isn't smoothness map that sharing with diffuse map alpha channel?

[marble rain]

It simply just sets a empty

[eternal garnet]

switching to RGB24 for instance where there is no possible A channel its fixed

[marble rain]

Eh. Don't use that lol

[sick relic]

Both standard and mochie have an option to use smoothness from diffuse alpha.

[eternal garnet]

testing purposes.

[eternal garnet]

NOW THAT might be something

[marble rain]

For instance. Using no alpha is nice to remove ehm seams basicly

[marble rain]

Between colors that contrast

[eternal garnet]

not even using specular

but

you are giving me an interesting idea

[marble rain]

What exactly is the problem

[eternal garnet]

The problem is if i use BC7 compression (or presumably anything with an alpha channel) i get transparency artifacts in the texture visible in masked environments such as your personal mirror

https://cdn.discordapp.com/attachments/512468176202891264/1131625233573490708/image.png

as shown here

you can see through the hair on bright spots and around the eyes

[marble rain]

Pooh I know that. Whm

[sick relic]

If rendering is opaque then it shouldn't be transparency. Rather it looks more like reflection.

[marble rain]

Forgot what's its called

[eternal garnet]

that is where BC7 apparently decided to include alpha

[marble rain]

It's not alpha

[snow knot]

Your metallic is set to 0 right. So it really shouldn't matter too much which workflow you pick here right?

[eternal garnet]

metallic is set to 1 as i'm using it for environment rim and specular reflections

but since i'm not generating specularity or metallic out of diffuse or alpha nor using a texture it should be uniform

[snow knot]

It definitely looks like a smoothness/metallic issue. You can kinda see the unity skybox reflection there by the eyes

[eternal garnet]

so if anything the transparency should be uniform across the body

[marble rain]

What you need to do rena. Have a clean material and apply each property one by one

Til you see the problem

[eternal garnet]

That's the window behind it

maybe this shows it a bit better

[snow knot]

It doesn't look in-line with the horizon I'm seeing to the right of the image

[eternal garnet]

the metal panel behind the mirror is coming through

[marble rain]

You have specular going on.

[eternal garnet]

its this metal wall coming through the body

but yes i have specular reflections enabled on my avatar

very very slightly

but again uniformly across the entire body

[marble rain]

Disable it

[eternal garnet]

so it shouldn't be only square spots

[marble rain]

And see

[eternal garnet]

already on it

still doing it

specularity is off

[marble rain]

Hmm

Looks like reflections

[eternal garnet]

its really not

[marble rain]

Or uv mapping issues

[eternal garnet]

again its the mipmapped metal panel behind mirror me

[marble rain]

Do a clean shader with just main texture

[eternal garnet]

what i was going for next

[marble rain]

And I can tell you mip map does not do that sadly.

[eternal garnet]

Eh, i can see a case or two, the mipmap alpha resize options in combination with some whacky shit.

[marble rain]

It's almost always a shader problem

Settings such

[eternal garnet]

yea but in case of switching shaders thats scorched earth for what is essentially a minor inconvenience that is almost exclusive to masked mirrors.

Plain new shader, same. Testing with standard now

standard shader

does the same

[sick relic]

Are you sure that there isn't any animation swapping the material? Try removing FX layer.

[eternal garnet]

good point

i added a shader switch recently

i was a little confuzzled seeing that the very clearly metallic look wasn't translating into VRChat

yea its gone

wow

https://imgur.com/yDeTdLT

it was just recently fixed...

So, conclusion: It IS an alpha issue, and BC7 DOES regenerate alpha into the texture into a completely opaque texture that doesn't feature an alpha channel to begin with and neither is unity set to generate one.

So despite that being fixed, can i still get a BC7 with forced alpha channel off?

https://tenor.com/view/cat-cutie-sad-sorry-puss-in-boots-gif-14112849

because you know stripping the alpha channel completely would save some VRAM

[kindred sluice]

Why are most the avatars marked as poorly optimized? How does VR chat measure it?

[eternal garnet]

Open your menu, select them -> avatar details

most likely due to too many polygons, followed by too much VRAM usage, followed by lights

[kindred sluice]

oh. so in avatar details it tells you the measurement of optimization.

[eternal garnet]

yes

for every category, it picks the worst and makes that your performance rank

[kindred sluice]

thx. it bothers me that most the avatars i saw are poorly optimized.

[eternal garnet]

welcome to user generated content

[kindred sluice]

it is just bad work then, you can make a great avatar with good optimiztion ?

[eternal garnet]

yes

[heavy knot]

VRChat performance limits are really bad to go off of a lot

[kindred sluice]

Time to learn how to make avatars.

[heavy knot]

you can have a excellent avatar bring you to 10 fps

[eternal garnet]

They are a bit iffy at times but they also dont include everything

[heavy knot]

basically make a VRAM zip bomb if you just use material swaps

or hell, coat your avatar in grabpasses

[kindred sluice]

Any good place to find good optimized avatars?

[eternal garnet]

they don't include things like state transitions, what animators do, blendshapes and some other values such as 2 skinned meshes + 8 static meshes are weirdly static numbers since you can use both together, not to mention its often recommended to split body and face due to blendshapes which immediately puts you at 2 skinned meshes leaving 0 room for anything else... and the polygon limit is quite extreme too INSTANTLY dropping you to very poor the moment you go over by a single polygon rather than having multiple stages like most other things

[heavy knot]

praise be the 1 million polygon potato avatar

that barely hits performance

[eternal garnet]

not to mention shaders are completely excempt from this, you can modify the standard shader to do an infinite while(true) loop and crash GPU's

[kindred sluice]

how to make a pretty face with best mathmatical ratio.. lmao

[eternal garnet]

yea and state toggles...

[heavy knot]

that method is pretty outdated for them anyway

load a bundle that unity doesn't like? goodbye

[eternal garnet]

i get that theoretically all meshes could be active at a time but why not have two values (one current polygon and one theoretical max polygon limit and generate a median out of that?)

[kindred sluice]

Doesnt blender have buildin optimization tool? Is AI a thing for 3d model building yet?

[eternal garnet]

Blender does have tools for optimization, tho probably not the 1 click solution you are looking for thats an extra plugin, AI modelling is a thing.

[kindred sluice]

With all the these poorly made avatars, i dont think one click optimization solution is good enough or coomon enough.

I am pretty sure AI optimization can solve 99 percent of the user errors.

[eternal garnet]

It requires some setup and most avatars are kitbashed

[kindred sluice]

whats kitbashed?

[eternal garnet]

jumbled together pieces

take legs from A, upper body from B, head from C etc

[kindred sluice]

ah..so unifying pieces together is not buildin.

[eternal garnet]

most avatars (anime ones that is) are just a couple bases mixed and matched with different pieces

[kindred sluice]

can blender tell which catagory of creature you are trying to build?

[bold wharf]

no

[eternal garnet]

no it doesnt matter

[bold wharf]

3d model is 3d model

[kindred sluice]

doesnt auto optimization reqcuire that?

[eternal garnet]

no

[kindred sluice]

So we are not there yet.

[eternal garnet]

auto optimization takes the model and decimates it with an algorythm according to parameters and modifers you give it, such as keep edges, mask specific parts to be weighted more so they are less decimated etc

i doubt an AI will ever be able to determine whats the best way to optimize an avatar

[bold wharf]

Theres also like... a lot... to optimising

Its not just reduce poly count

[eternal garnet]

its called intuition, something the AI doesn't have, something like "this is an elbow, it needs to be a bit more detailed to allow for better deformation" the AI can't tell that its an elbow unless you were able to tell it somehow since its just 3d coordinates, thousands and millions of them

well yes but Grimms was talking about AI modelling

so the main optimization here is polycount, followed by material count and blendshapes

[bold wharf]

When you get down to the real nitty gritty parts as well you start coming to tradeoffs where you need to make subjective choices to whether sacrificing a feature/quality is worth the tradeoff for performance loss. Thats not something an AI can do for you

Easy example being say you want to have nice flowly long hair, but to make that look decent it might take a lot of polys. Do you sacrifice having long hair to keep the performance or do you say "fuck it" and go for the long hair anyways (this is just a quick hypothetical example)

[heavy knot]

it's a horrible design choice to put poly count as the first metric you're met with to users

as unsuspecting people are now going to think polygons is what causes lag

[kindred sluice]

There must be a lot illusion tricks for high effect and low poly, right?

Mostly use light?

[eternal garnet]

it is most commonly one of the biggest factors in general and contributes to lag in many ways.

while putting it as the BE ALL END ALL ultimate reason for lag is wrong, it is very okay to make it one of the main categories to look out for

since it plays into multiple things and compounts with them depending on other things

a model with many polygons and blendshapes most likely moves a lot more polygons per blendshape, so polygons once again indirectly influence blendshape performance

just as example

[kindred sluice]

arent there a lot illusion tricks? for max effect and low poly?

[eternal garnet]

normal maps yes

[kindred sluice]

ah, so use pic instead of 3d poly

[bold wharf]

I mean sure but you can't fake what isn't there.

[kindred sluice]

yeah.. i get that.

[eternal garnet]

normal maps, detail maps, ambient occlusion maps, specular and metallic maps, parallax mapping

[eternal garnet]

You very much can.

[bold wharf]

Normals maps are good for detail but if you go low poly you'll still have a blocky silhouette

[kindred sluice]

Human eyes are easily decieved..

gotta learn some illuison trick from human eye study. lmao.. max avatar studying.

[bold wharf]

Yeah but you still need the geometry there in the first place is what I'm getting at. y'feel?

[kindred sluice]

kekw

[eternal garnet]

do you think that detail is really there?

it really absolutely isnt

[kindred sluice]

so many furry.. bro. i hate furry. makes want to hug them. lol

what details?

you mean the fingernail?

ah

the fur

nvm

[bold wharf]

I think we might be on different pages here Renamon.

[kindred sluice]

need a better pic for the fur.

[eternal garnet]

i'm talking about the body normals, aka the "muscles" the visible bones/bumps

[kindred sluice]

ah.. i see

[bold wharf]

I'm trying to say you can't fake details such as those hair/fur tufts on your tail and legs.

[eternal garnet]

well yea

[kindred sluice]

yeah. color change maeks it look quite real.

[bold wharf]

Yeah so we were on different pages lol

[eternal garnet]

not really

there are things that simply need to have the basic geometry

but beyond that you can fake a lot

entire fur

[bold wharf]

Yeah but thats what I was saying

[eternal garnet]

thats what fur shaders are for

[bold wharf]

You cant fake what you don't have. If you don't enough basic geometry there what good are normals and whatnot gonna do.

[kindred sluice]

any good tricks for hairs?

like meshing them together and draw white line between them?

[eternal garnet]

hair depends on what kind of hair you want to make

realistic hair can be made in many different approaches depending on the hairstyle, then theres simple flat hair (cartoon hair) and theres also as shown above the basic "tuff" version

[kindred sluice]

A lot anime characters have really good looking hair, anime drawing could be used for avatar building a lot.

https://tenor.com/view/the-eminence-in-shadow-the-eminence-of-shadow-shadow-garden-kage-no-jitsuryokusha-ni-naritakute-kagenou-gif-27024397

[bold wharf]

I'd reckon probably about a 3rd of the avatar market is all based of that style (eastern market)

It is very common

[kindred sluice]

i am currently look for this anime character avatar. Do you know where i can find it, good sir?

This anime is called eminence in the shadow

[bold wharf]

booth.pm has the largest collect of anime styled characters. I do not know if that specific character has been adapted into 3d and is purchasable however.

Unfortunately there is a very huge lack of male characters in comparison to female ones

[kindred sluice]

thank you.. yeah. female ones are everywhere which makes you think.. why man likes being a female so much in VR.. hmm lol

[marble rain]

Not exactly. Bump maps normal maps. Height and detail maps can make low poly look like high. By using a high poly version of the same product and build from that onto low

[bold wharf]

So if I'm not mistaken height maps work by actually raising and adding real 3d detail. But is that more performant than extra geometry?

[marble rain]

It is

[bold wharf]

You got numbers?

Or source

[marble rain]

You can gain the same effect as from high poly

[bold wharf]

I mean no disrespect just want to be sure about these things

[marble rain]

Because your poly are lower. And lower complexity

It's a common thing to do in game dev btw

[ivory sluice]

Height maps can work as either parallax maps, which adds a 3D height effect stereoscopy-wise, or can be converted to normal maps by comparing the difference of samples of that texture shifted on the uv cooridnates.

Most times it's both in one

And it is more performant than 3d geometry, at the cost that you can't detect when it reached the model's edges, so it can be tricky for models with unpredictable shapes

[bold wharf]

Yeah so still not really applicable to what I've been trying to say I guess.

I've not been saying that you should go and use 3d modelled fishnets here

I'll go back to my hair example, sure you can use a normal map or whatever to add detail to it, but that normal map won't magically give you a long ponytail or something

[ivory sluice]

Yea obvs you have to model it

[bold wharf]

This is all I've been trying to say this entire time 😭

[ivory sluice]

Idk i just joined

[bold wharf]

Yeah you good wasn't trying to bash you or anything I just feel nobody was on the same page as me this entire time

[flat sage]

I've got a merman avatar I'm planning on making in PMX and was wondering whether non humanoid rigs are allowed on the quest version?

[tawdry drift]

i dont see why not

but youll have to remake all of the animations if you're doing non humanoid rig

i dont see why a mermaid would need to have a non humanoid rig though. just use the upper half of the rig as normal and weight the mer tail to one of the legs only

[flat sage]

Ah cheers for the tips I'm not too savvy in how the rigging works 🙏

[buoyant holly]

yeah basically you can have the humanoid bones weight painted to nothing and still have it be a valid humanoid Avatar

like any snake avatars would basically be a Unity humanoid with the legs weight painted to nothing and the tail being physics bones

[flat sage]

How would I do this weight painting? I'm presuming it's done in blender as I've heard about it in my very minimal experience in the program :>

[buoyant holly]

yes weight painting occurs in blender

[flat sage]

Alr so would I have to create some leg bones then have them just not weight painted at all?

Sorry if I'm being really stupid

[buoyant holly]

yes

[flat sage]

Alr thank you!

[buoyant holly]

https://www.youtube.com/watch?v=JFBQeNON64Q Tupper Avatar optimization panel is now on YouTube

[rain matrix]

Bro was walking on eggshells the entire presentation 😭

[ivory sluice]

Tbf it's ironic how he made such a good presentation on optimisation when that same convention made the unintentionally laggiest avatar shader my gpu had to ever render

[rain matrix]

💀

[buoyant holly]

the videos up on YouTube now so it can act more as a standalone piece

[buoyant holly]

how so out of curiosity?

[rain matrix]

Tupper has a lot of eyes on him and he has to be very careful about the things he says and I feel like he is very aware of that

One piece of misinformation from him can cause huge ripples

During the panel he never really went that deep into anything specific

[buoyant holly]

runtime considerations would probably prevent going to super deep into stuff

[buoyant holly]

yep as Tupper has the unfortunate job of basically being the mascot of VR chat

[ivory sluice]

Yea kinda

More of a mascot it's that most of the users think that he is the creator of everything related to vrchat

[slate crypt]

How performance intensive would 20 contact senders be if only one is enabled at any given time?

[somber raven]

You should be fine. Although since it’s 20 Contacts, it will place you at Medium Rank no matter if it’s enabled or not.

There is a hard-cap of 256 Total Phys Bone and Contact Components per Avatar being implemented sometime soon, which fixes crashes related to excessive contacts colliding at once.

[lime hemlock]

how much should i reduce this to be an avatar?

[buoyant holly]

to not be very poor that number needs to be 70000 or less

[lime hemlock]

oh then i have room ig

gotta finish the hair, have like 50k+ poly to spend if you're right lol

[buoyant holly]

https://creators.vrchat.com/avatars/avatar-performance-ranking-system/

[clear raft]

Would anyone be willing to show me the basics in make an avi? (Didn’t know where to put this mb)

[obsidian lake]

basics really depends on how much you want to do from scratch

there are definitely basics to the unity side of things, but in terms of where your model comes from, it can be as easy as buying someone else's model, or you could make one completely from scratch, or somewhere in between

[proven dock]

https://www.youtube.com/watch?v=bSwMz4WcajQ&list=PLt8qiEFqNHFkPqpKLnVRx-G78mGvjPbFt https://www.youtube.com/watch?v=JFBQeNON64Q

[clear raft]

Oohhhh

Okay Ty!!

[snow knot]

iirc this already got rolled into one of the recent updates. I think it was the last one w/ the custom emojis

[open pagoda]

Working on my friend's avatar, and he wants a million-billion outfits, toggles and everything else all on one avatar. I tell him "no, those should be separate uploads", but he won't have it- he'll "only wear it in 1:1s", so he says... And I see him pulling it out in these lobbies of 25 people.

I decided to invent a new performance rank for him.

Slapping this on the avatar thumbnail so he's forced to confront what he's doing every time he swaps into it.

[balmy escarp]

you can't atlas a material texture with animations in vrchat, right?

I have it set up for my vtuber model, but I couldn't figure out how to do it with animations

like I mean for creating texture toggles for my eyes, face, etc

you can't just use a single material for all of the eye textures?

[spring sun]

you can

sprite sheets

[spring sun]

sprite sheet with animations that offset the texture

and so on

[balmy escarp]

ooo, I'll check it out, thanks!

I'll hold onto that in case I need to optimize my avatar further lol I'm not even to Medium yet, but just in case

should be useful for my bunny girl avatar, which already sits at Very Poor because I wasn't very optimal in making the clothes for it

[lone tiger]

That's most of Chinese and Taiwanese Booth personal avatars in a nutshell.

[opal garden]

is there a way to let a separate Animator Controller get Synched Variabled? im trying to Optimize my Animator by Moving the Heavy Animations to a Separate Animator that is disabled while the object is not used.

[lone tiger]

It was possible in the past, but now it no longer works.

[proven dock]

how many mat slots is it?

[open pagoda]

It was somewhere on the order of magnitude of 20-30, haha.

[pliant nymph]

any reccomendations to be able to shave off roughly .4 mb off of an avi thats .35 mb too big for quest? kinda unsure of what to do from here. odd thing is that i was able to upload the same avatar before and it was under the limit

[tawdry drift]

.4mb hmmm

maybe you could yoink a blendshape, see if theres some geometry youll never see that you can remove, dissolve like 2 edge loops lol

thats such a small amount im not sure what itd exactly quantify to

reduce the size ofa normal map, maybe a matcap?

[pliant nymph]

i was able to squeeze .4 out of it by reducing the size of some of the textures that i dont generally use (materal swaps n stuff)

down from 1k x 1k to 32x32

ty for the suggestion if i need to compress it further i shall do that c:

was able to actually get it down to 9.2 mb (when compressed) :D

[tawdry drift]

nice, epicc 🙏

[opal garden]

-.- dam this makes things complicated... to understand and master the unity animator you probably need 5 Master Degrees and 3 Doctor Titles...

[proven dock]

30 material slots isn't even very poor though

[marble rain]

30 materials is far to many for an avatar.

[proven dock]

On that note does anyone have a tutorial on how to convert a mesh toggle to an identically-acting blendshape toggle? (or even better a tool or Blender script that will do the hard part)

(assuming moderate familiarity with Blender and Unity)

[eternal garnet]

Why would you do that? A blendshape toggle is far far worse for performance and costs a lot of VRAM.

If you want to toggle something and want to keep the SDK performance rating happy, use shader toggles, most shaders can turn off "rendering" by setting them invisible. Mochi for instance has it in Render Settings. It's not perfect and will still incur a draw call but it will skip the entire shader.

[proven dock]

biggest problem with shaders is if I put it on a public avi anyone who isn't trusted user will have them broken by default and not know it

[eternal garnet]

Most people have everything off anyway

and if you default the avatar to something like the first outfit for instance, you should be fine

also custom animations is a thing too, people might have that disabled, which will break your stuff too

Again most people have everything off anyway, that or have everything enabled, sadly there is hardly any inbetween

I've always had all avatars enabled with all features but i've recently grown a lot more aggressive with global muting avatars that eat a lot of VRAM

keeping VRAM usage low is a lot more important than most other things atm, considering that most other things in the ranking currently are quite heavily optimized

and VRAM is usually heavily limited and will instantly turn your shit into a slideshow the moment you start running out

[buoyant holly]

people can turn off animations on avatars which would break any toggle so you might as well go with the most vram efficient options

[eternal garnet]

when you switch shader off, switch material to an empty shader too, thats going to reduce the active memory usage

[marble rain]

can also use one bone per equipment. with no weight paint or anything, making them into equipment that can be disabled and such. however i dont know if this is much worse then say something like UV discard.

[hushed bramble]

.

[limpid cargo]

well this is the definition of mildly infuriating

[tawdry drift]

oh nooo, 34 poly, oh noooo you're gonna crash people >:( >:(

the horror

[bronze lantern]

Is it better to have 35k polys in one mesh or have two separate mesh renderers with less polys?

[ivory sluice]

In general 35k polys in one mesh is better

Things like an excessive amount of blendshapes could make it worse tho

[heady smelt]

Should only become a problem when a single mesh is over like 70-75k triangles

Some threshold around there

Forgot what it is exactly but it's something

Someone else here knows

[ivory sluice]

After going to more than 65535 polygons, the indexes will have to use 32 bits, which increases the mesh data memory. I don't think it's a performance impact since nowadays gpus don't process in less than 32 bits

[heady smelt]

i feel like the poly restrictions on avatars is strict asf

even when im in public worlds with no safety settings my performance doesnt drop much at all and you alr know some of those avatars have like 300k polys lmao

[heady smelt]

broo I need to combine my materials lmao I have 49

[snow knot]

I think it's mostly a filesize improvement. I know I knocked a bit of filesize off one of my avis by splitting the body and head into seperate meshes

[ivory sluice]

yeah it's obvious modern gpus process vertices blazingly fast, vrchat probs has this maximum as a worst case scenario, because shaders are responsible of vertex performance, and these cannot be measured.
So the whole scene (world + all avatars on the instance) could sum a total of 1M polygons (as an extreme case because for me that's a terrible poly count), but at the moment any avatar or part of the world turns on a realtime pixel light, that poly count doubles, if it has shadows then it adds more, the multipliers for polygons to render add as follows:

• Realtime pixel light : +1
• Realtime pixel light with shadows (spot or directional) : +2
• Realtime pixel light with shadows (point) : +7
• Depth buffer, which gets enabled once when any realtime light has shadows : +1

and the multiplier also counts for any extra passes or geometry shaders, for example an outline, fur shader, or the furality sylva shader which for some reason it needs to render a second time for the effects

so depending on what is and isn't on the scene, your gpu might be rendering a lot more times the poly count that your avatar says it has

and i think 70k is good enough for an avatar, i got to make an avatar with a fton of mechanisms and it got to 50k or something

[snow knot]

Also mirrors and cameras will multiply your rendering time iirc

[somber raven]

Cameras especially will multiply your Rendering Time because you are outputting to a Render Texture.

[ivory sluice]

yes mirrors and cameras will double that multiplicator i forgot to mention it, the multiplier can be slightly reduced by disabling pixel lights on the mirror

[drowsy fjord]

can I like somehow keep my Phys bones without completely destroying the performance rankings? my avatar by itself is like "good" on the performance scale but... the moment I add Phys bones for like hair and skirt the performance ranking drops down to Poor...

8* components is not much to work with :/

(just for reference I have 31 components)