Is VRChat really preloading all textures / materials contained in an avatar in VRAM
even if they are not inside a material slot and used on the avatar at a given time ?
From what I understand in your new ranking you’re going to sum all the texture sizes that are in the avatar package and check blindly if that sum goes over the limits ?
I try to optimized my avatars a lot and stay under medium perf,
also at a given time my avatar is using only one or 2 materials with its associated textures. Let’s say 2x2k textures as an example.
So it’s still super optimized imho.
and it’s normally not using VRAM when they are not loaded ?
But as my avatar packs multiple materials with different textures that I swap in real time
for different effects, that would rank it poor or verypoor now because the sum of all the individual unused textures at a given time would still go over the new ranking limit ??
Well right, but I specifically mean adding identical bones in the hierarchy. My understanding was that this acts nearly like constraints since it needs to calculate the position of the extra bones and items added in the armature just like a parent constraints. Unless adding objects in the armature is cheaper than parent constraints?
Also yes, I do agree that adding clothes in Blender is relatively trivial. But many users also don’t want to get near Blender, and other’s aren’t aware of the tools to properly swap meshes so it uses the armatures correctly. Not to mention, many of those scripts may not work or may be broken in some ways. Was trying to help someone with clothes that were made for their avatar and both versions of the mesh swap script did not work.
As I understand it: It will load all textures (and meshes) into VRAM, and then swap out textures that are not active if the VRAM gets full. Though this is to not be confused with mipmap streaming, which Fax said VRChat does not have (even though the SDK does enforce mipmaps if they are enabled for the material/texture, which is good if they do add it in the future).
Yes, because they still get loaded anyways. This is more how Unity works. Even if your avatar is decent, if there are too many active avatars around you, then your FPS will drop down to about 2 FPS since it will be actively swapping textures in and out of VRAM since the total amount of active textures will exceed the amount of VRAM you have. The proper solve for this is for VRC to enable mipmap streaming so it can more reliably swap out different resolutions of textures at runtime, to more efficiently use VRAM.
If your avatars are going to be very poor, then you already were causing problems with VRAM performance. Just because you did things to optimize doesn’t mean you have an avatar that is actually optimized to run well in the traditional context of video game optimization in a controlled environment.
To note: Just because you Atlas something doesn’t inherently make it better or worse. Though if anything, Atlasing may cause more VRAM usage. For example, if you have 2 or 3 1K textures and you Atlas them into a single 2K texture, you will have wasted space. If these are also textures that are swapping rather than being shown at the same time, it makes more sense to just swap them on the material without atlasing.
The entire point of atlasing is so you can use less materials, not necessarily to reduce VRAM usage. Less materials means less draw calls and less overall rendering. If you have, say, one 1K texture, and a dozen 256 textures each with their own material, you would have 13 materials on X number of meshes.
If you were able to merge all those meshes and then Atlas all those materials together into a single materials, then you go from 13 materials and 13 meshes, to 1 material and 1 mesh. This is an extreme example, but this is part of the reason that maps exist. They allow you to have various roughness, metallic, emissive, etc. This is why maps are powerful, especially if you can rescale them nicely to a lower resolution. This is also why good UV mapping is important. It’s why tiled, repeating textures are so awesome, because you can get away with a ton of detail with even just a 512 or 256 texutre.
Even normal bones when imported into Unity are just GameObjects with weight values for a skinned mesh, adding more of them is exactly like adding more bones in Blender, except in Unity.
Might seem like they’re different, but they’re really not.
So it’s something in between:
they are indeed preloaded, but marked as inactive and so can be freed up
at any time if something needs the VRAM.
While I understand it can cause a slight overhead as they are unloaded, I think it’s really minimal.
Compared to an avatar with 10x4k material slots and clothing or 8K textures that would be constantly on.
While the new system will clearly help to detect such avatar (and I’m glad it does)
Mine would kinda be treated the same , while it’s not impacting performance the same.
So 90% of my unused texture would get (pre)loaded, then freed once to make space as soon as others need it.
That still seems super weird to me that unity needs to preload everything contained inside a package.
How do big games with 50GB textures on the HD do ?
They don’t preload the 50GB in VRAM afaik ?
Can’t vrchat do modular loading of what’s inside the package when it is really needed / used ?
or is it because the smaller item unity can preload is the entire package or something ?
What will most likely happen here is that people will have to split their big avatars into multiple vrc avatars, and do multiple avatar swaps while being in a world …
Resulting in a even worse end user performance imho
Even if avatar loading has been more optimized in vrchat, it still generates lags.
Cool! Technical Details are always really appreciated
One thing i’ve wondered:
Are there currently any plans to change the Collider Capsule for players?
It would be really nice if it was attached to the hip instead of the head, for all the FBT users who want to lean over a small object, or look down a cliff.
Sorry for the late/second post on the same dev update but I had an idea when I awoke this morning that could be useful for some less experienced Unity users.
Would it be possible to implement a “guest key” or some other method of commissioning someone to upload an avatar for you, but where you give them a private key or code that allows them to upload the avatar to your own account without granting them full access?
For obvious reasons, I’ve never shared my login info with anybody. But it would be great if a friend who created an avatar could upload a version of their Avi to my account without having to give me all the files, and without having to mark the avatar as public and potentially having many copies running about if I make an error with clone settings.
I think it would be a great way to share private avatars within VRChat without giving away all the source material all while keeping user accounts secure. (Nobody would ever need to feel pressured to share their login info)
Just a thought for SDK updates down the line. Thanks! Have a great Friday and subsequent weekend!!!
you have a maximum of 8 material to reach good performance ranking, you can definitly finds way with atlasing. Not everything need to be 4k , you can lower rez of texture that have little to no detail.
So as a question for skirts. I currently use rotation constraints + phys bones, but would it be actually better to use colliders instead of the constraints? Possibly looking at one plane collider per skirt bone chain.
Cause it was that constraints are supposedly cheaper than dynamic bone collission. Is it now that phys bone collission is cheaper than constraints?
if you are using more than 1 set of 4k textures (diffuse, normal, mat) per avatar then you should downscale. no one needs a 4k texture just for a single part like hair or shoes.
And even on my 1 mesh 1 mat models that use 4k for diffuse I usually still keep the normals and other maps at 2k or less.
Do you have any technical breakdown or Unity docs that explain this further? I guess I’m just not understanding, because what you’re implying is that binding an object to a bone in the armature is cheaper than using a parent constraint. Do I have this right?
Most games use mipmap streaming. But most games also have very optimized models and textures, to where even one that looks great might be good or medium rated on VRChat with low texture memory.
VRChat does not have mipmap streaming enabled, according to Fax. You can upvote this canny here asking them to add mipmap streaming.
I recommend upvoting the canny for that. @docteh did already link one, but this one has more upvotes.
Yes it’s cheaper, same cost as adding extra bones.
As for documentation, no.
Try unpacking a model however, you’ll notice you can move bones to other objects willy nilly, the part of the mesh will still follow it, but it’s really just a game object with mesh weight data.
Wow that’s a lot for this Dev Update!
I’m a bit concerned for the hard work that I could put in my avatars to be a little bit more optmized, but I really don’t want to decreare the “fancyness” and visual quality of my creations.
I totally feel this. But at the end of the day, this is a social game with many users with many setups. In my opinion, I am willing to sacrifice some quality if it means people can more reliably see my avatar at higher frame rates. “10 foot paintjob” works pretty well in this instance.
(says me, trying to max out both optimization and quality in my world )
Big game load texture in a scene at given moment. Vrchat scenes are basically worlds. Those games are using multiple time the same texture at multiple location because they can do that, you can’t do that with avatar because everyone’s avatar is different.
Not every textures need to be 4k, you can reduce some texture that have little to no detail.
)