VaeDecodeStructureTrellis2 - ComfyUI Built-in Node Documentation - ComfyUI

This node converts Trellis structure latent samples into a 3D voxel grid using a VAE’s structure decoder. It reads only the first 8 channels of the latent, reconstructs the voxel occupancy, and adjusts the output resolution to either 32 or 64 as requested.

Inputs

Parameter	Description	Data Type	Required	Range
`samples`	The latent representation of the structure to decode. Only the first 8 channels of the latent are used for decoding.	LATENT	Yes	-
`vae`	The VAE whose structure decoder converts the latent into a voxel grid. Decoding is performed in batches.	VAE	Yes	-
`resolution`	The target spatial resolution of the output voxel grid (default: “32”). If the decoded grid has a different resolution, it is downsampled to match.	COMBO	Yes	”32" "64”

Note: When the decoded voxel grid resolution differs from the selected resolution, the grid is downsampled using 3D max pooling to the requested size.

Outputs

Output Name	Description	Data Type
`voxel`	A binary voxel occupancy grid as a float tensor with shape [batch, depth, height, width]. Values are 1.0 for occupied voxels and 0.0 for empty voxels.	VOXEL

This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub

Source fingerprint (SHA-256): 37764ef7351b3619d4cddb57b11d9a0da24dadeedc0fc0f70d089038d37e03b0