Pixal3D Shape VAE (CRT AutoDL)
The geometry half of Pixal3D's two-stage brain
- VAE
Pixal3D doesn't generate a finished 3D model in one shot. It runs a two-stage pipeline inherited from the TRELLIS.2 family: first the shape stage decides what the object actually is - its silhouette, volume, where the surface lives in space - and then the texture stage paints it. The shape VAE is the part that encodes and decodes that geometry latent. If the model node is the brain, this is the half of it that thinks in voxels.
CRTAutoDLPixal3DShapeVAE downloads trellis_2_shape_vae_bf16.safetensors into ComfyUI/models/vae/ on first run and loads it as a proper VAE. No inputs, one output, zero configuration. It's the purest version of the pack's AutoDL pattern: check if the file exists, download it if it doesn't, load it, hand you a VAE.
The one output
VAE - this goes to the shape side of the pipeline: the native VaeDecodeShapeTrellis node decodes the sampled shape latent into the actual geometry. (The companion CRTAutoDLPixal3DTextureVAE does the analogous job for the texture side via VaeDecodeTextureTrellis.) Both feed from the same two-stage flow: Pixal3DConditioning → shape stage → texture stage → decode.
Install
The whole pack installs the same way:
cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes
pip install -r requirements.txt
Restart, or use ComfyUI Manager and search CRT-Nodes. Because these nodes delegate to core comfy_extras nodes, you need a current ComfyUI - if VaeDecodeShapeTrellis doesn't exist in your node search, update ComfyUI first.
Gotchas
- One-time download on first run - a progress bar appears in the console labeled
CRT AutoDL. The file lands inmodels/vae/and is then reusable by ComfyUI's own VAE loader, so it's not wasted space specific to this pack. - Don't mix it up with a normal image VAE. This is a shape VAE - it operates on the 3D structure latent, not on pixel space. Wiring it into a regular image VAE decode will fail, because the socket type it feeds (the Trellis2 decode nodes) is different from the one an image
VAEDecodeexpects. Keep it on the Pixal3D side of the graph. - The "red node / NaN" trap. If you update CRT-Nodes and existing graphs show a red node or
NaNwidgets, right-click → Fix node (recreate). The README calls this out; it's socket changes between versions, not corruption.
Honestly, this is a utility node with a small job, and that's fine. Its whole value proposition is "don't make me find and download trellis_2_shape_vae_bf16.safetensors myself" - plus loading it with the right bf16 settings so the decode actually works. If you're assembling the Pixal3D workflow for the first time, you'll want this node, its texture twin, the model selector, the DINOv3 CLIP vision, a MoGe node for FOV, and a background-removal step for clean input. The pack gives you all of them under CRT/AutoDL/Pixal3D, and this is the one that carries the geometry.
Inputs (0)
No inputs
Outputs (1)
| Name | Type | Description |
|---|---|---|
| VAE | VAE | — |