I was going to say that this mipmap is not as good as forcing on anisotropic 16x boosting in a real 3D environment.
Then I experimented with anisotropic sampling to my own project and was silenced on the spot…
Hmmm… That’s a nice boost.
The PC’s texture units are adaptive, resulting in varying performance in different scenes, unlike mobile where different texture filter settings can have a significant impact on performance (too fixed and dead).
With the general in-game stylised shader coupled with the lack of depth culling resulting in a high pixel shader load, forcing anisotropic 16x is too much of a performance hit…
(Especially with poor UV and other asset design, which makes it better to enhance those parts instead of using texture filtering anisotropic 16x).
As for why it generally doesn’t affect 3A games at the moment? The screen design and scenes are relatively flat, and the calculations that can be used replace or reduce sampling, and the sampling process can be implemented with the help of compute units, unlike Unity which is relatively simple to run on a set of drawing APIs…
(In fact, I rather suspect that there is actually a set of texture management methods inside, and the actual sampling is forced to clamp down, etc.)
The driver can be set to cause the texture unit to downsample (tri-linear to bi-linear and the hardware to realise the upper level automatically adjusts down to meet the demand can be).
I’d recommend fixing the mipmap stream problem first though, that’s more important…
It might be possible to implement an SDK override for the Unity engine to adjust the anisotropy to reduce the growth of the texture unit sampling computation cycle, thus reducing the pixel shader load and time consumption.
Although most of VRchat may only have trilinear and bilinear, it basically doesn’t affect the performance…
You can save performance or trade it for enhancements such as texture blurring and moire on bulletin boards in some worlds…
But also ask the official avatar on its own calculation will affect the efficiency of rendering, the more mesh the more obvious, but even without mesh avatar compared to pure project export is still 10~20% slower, I do not know if there is any modification to the Render or have the camera forget to correct?