yeah , first thing i did was check the colorspace since both you and Akella got screwed by it 
new test
16 images camera rotates from -12 to 12 deg , animation has linear interpolation , vector png image is normalised + has raw view transform
1 Like
the watch’s CompressedArrayTexture is able to hold 200 textures with1024x1024 res in 2.18mb ?!
will experiment with it next 
1 Like
@vis_prime Yeah, + mip levels it looks like 
it’s a staggering difference! Also keen to dive into how these would be packed, if you find any good pointers let us know 
tested ktx compressed texture array test
settings used
ktx_command = [
"ktx", "create",
"--format", "R8G8B8_UNORM",
"--layers", str(len(files)), # total num of images
"--encode", "basis-lz", # Encoding method
"--clevel", 1, # compression quality
"--qlevel", 128, #quality
"--assign-oetf", 'linear',
]
(then append all the input image paths and the output ktx path)
Input : 180 color pngs and 180 motion vec pngs ;1024x1024res totaling 333MB
after ktx compression , output was
4.7mb color & 2.6MB vec ktx2 textures
observations:
- if basis is not used the ktx texture is around 200mb
- blocky shapes work best (that’s why i chose the phone model)
- textures are upside down , so had to account for that in the vertex shader and displacement calc
- compression artifacts are easily visible.
- renderer.info.memory.textures says there’s just two textures in use even though one texture contains 180 images !
3 Likes
advanced solution to fix distortion on hard edges
might be useful for some one
https://x.com/ValigurskyM/status/1938613357625376809
It’s possible just Disocclusion with color clamping would work here: https://elopezr.com/temporal-aa-and-the-quest-for-the-holy-trail/… That’s what’s used here: https://playcanvas.vercel.app/#/graphics/taa A TAA without motion vectors. Color clamping solves the moving cube well enough.
2 Likes