Trellis2ShapeStage - ComfyUI Built-in Node Documentation

This node sets up the first shape-generation sampling pass of the Trellis2 pipeline. It takes the dense structure voxel produced by VaeDecodeStructureTrellis2, extracts the sparse coordinates of the filled voxels, creates an empty sparse latent, and attaches the sampling metadata to the conditioning so the model can read it during sampling. For the second shape pass after upsampling, use Trellis2UpsampleStage instead, which combines the cascade and the second-pass stage setup.

Inputs

Parameter	Description	Data Type	Required	Range
`positive`	The positive conditioning to prepare for the shape stage. May be a standard Trellis2 conditioning or a Pixal3D conditioning that supplies a projection feature pack; when projection features are present, they are computed for the selected stage and attached to the output conditioning.	CONDITIONING	Yes	Any Trellis2 or Pixal3D conditioning
`negative`	The negative conditioning to prepare for the shape stage. The same shape-stage metadata is attached to it as to the positive conditioning.	CONDITIONING	Yes	Any Trellis2 or Pixal3D conditioning
`voxel`	Dense structure voxel from VaeDecodeStructureTrellis2.	VOXEL	Yes	Any voxel grid; the grid resolution (voxels per axis) selects the pipeline stage

Notes

The voxel grid resolution selects the pipeline stage: a resolution of 32 or lower uses the shape_generation_512 mode with the shape_512 stage; a resolution greater than 32 uses the shape_generation mode with the shape_1024 stage.
The voxel must contain at least one filled voxel; an empty voxel raises an error. The batch indices derived from the voxel must be non-negative and contiguous.
When the positive conditioning contains a proj_feat_pack (as supplied by Pixal3D conditioning), projection features are computed for the selected stage and the output latent’s model frame is set to y_up. Otherwise, no projection features are attached and the model frame is set to z_up.

Outputs

Output Name	Description	Data Type
`positive`	The positive conditioning with the shape-stage metadata attached: generation mode, sparse coordinates, per-batch coordinate counts, and projection features when the source conditioning provides them.	CONDITIONING
`negative`	The negative conditioning with the same shape-stage metadata attached.	CONDITIONING
`latent`	An empty sparse latent tensor (shape: batch size, 32, token count, 1) together with the extracted sparse coordinates, per-batch coordinate counts, coordinate resolution, the type marker `trellis2`, and the model frame orientation.	LATENT

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): 7dbee8a5b6ef7111f07def4dbe1cc4908533e00ffcb775f5a284099360c7eed3

​Inputs

​Notes

​Outputs

Inputs

Notes

Outputs