VQField (@vqfields) header
VQField (@vqfields) profile picture

VQField

@vqfields

Neural Compression for 3D Volumetric Fields

Joined June 2026
  • Tweets9
  • Following7
  • Followers4
6Media
Photo by VQField (@vqfields)Photo by VQField (@vqfields)Photo by VQField (@vqfields)Photo by VQField (@vqfields)Photo by VQField (@vqfields)Photo by VQField (@vqfields)
1
4
911
👀
Text "Compressed data, Rendered on Demand" with the render of a cloud. VQField promotion image.

ALT Text "Compressed data, Rendered on Demand" with the render of a cloud. VQField promotion image.

1
1
23
It only takes a few nanoseconds to sample values. This allows a few cool things, for example, a renderer can sample the value without having to decompress the volume. Needless to say this would also work in simulations !
Enzo Crema (@ZephirFXx) profile pictureEnzo Crema@ZephirFXxSep 7
VQFields can now be rendered without ever be decompressed in Karma ! The 100x file compression is now also 100x RAM compression at render time :) #houdini #karma #vfx
Render of a compressed smoke plume in Karma, Houdini.

ALT Render of a compressed smoke plume in Karma, Houdini.

2
3
89
Volumetric simulations repeat a lot of information over time. When motion is slow, VQField can reuse this temporal coherence. Sequence compression can reach 200:1.
Image from VQField (@vqfields)
1
1
3
26
Velocity is not only simulation data. It can also guide prediction. Advect the previous field, predict the next frame, and encode only the residual.
Image from VQField (@vqfields)
1
2
24
Now with parametrable compression ratio !
Image from VQField (@vqfields)
1
6
vqfield.com
VQField — Random-access compression for OpenVDB

VQField — Random-access compression for OpenVDB

Compress OpenVDB volumes and sequences into files that stay readable while compressed. Explore measured quality, Houdini and Karma CPU integration, the Studio CLI and C SDK.

vqfield.com
1
10
1
2
41
VQField doesn't try to make 1 model fit all use cases. Since VQ-VAE models are so tiny, it's easy to load them in memory and dispatch them accordingly. One could compress SDF fields while the other is working on compressing Vector fields.
Enzo Crema (@ZephirFXx) profile pictureEnzo Crema@ZephirFXxJun 1
Replying to @ZephirFXx
VQ/codebooks are powerful: compact indices, reusable local structure, cheap decode. But the codebook is also the bottleneck. If the right pattern is not in it, the codec needs residuals, fallbacks, or accepts error.
2
24
Load more