fernandotonon/QtMeshEditor-trellis-ss-encoder-onnx

TRELLIS sparse-structure encoder: ONNX export (shape guide)

ONNX export of the sparse-structure VAE encoder ss_enc_conv3d_16l8_fp16 from microsoft/TRELLIS-image-large (MIT code and weights). All credit for the original weights goes to Microsoft.

TRELLIS.2 reuses TRELLIS-1's sparse-structure VAE, and the native trellis.cpp runtime ships only its decoder. QtMeshEditor runs this encoder to turn a rough blockout (a house, a vehicle, a body) into a sparse-structure latent, and starts TRELLIS.2's sparse-structure flow from it (training-free blockout control in the style of SpaceControl). The result keeps the image's look while following the guide's size, placement and volume: qtmesh generate3d image.png --backend trellis2 --guide house --guide-strength 0.5, the "Shape guide" row of the GUI, or the MCP guide argument.

The file QtMeshEditor downloads at runtime lives in the shared fernandotonon/QtMeshEditor-models repo under trellis2/. This repo is the standalone model card + mirror.

Examples

The guide steers the generation's volume; the guide strength goes from "image only" (0) to "follow the guide" (0.9).

A spring primitive guiding a coiled snake. Same FLUX.2 [klein] image and seed at every strength; the large one is strength 0.9 with the texture baked onto the spring itself (it keeps the low-poly topology).

A body as the guide. The armour comes from the image; the build and pose come from the guide (a MetaHuman body, and a heavier body made with QtMeshEditor's MHR character generator). Around 0.6 the outfit stays complete (cape included) while the build changes; at 0.8 the guide's pose takes over and loose cloth is dropped.

Files

file	role
ss_encoder.onnx	the encoder, fp32, opset 17, ~237 MB
LICENSE-TRELLIS.txt	MIT (Microsoft TRELLIS)

SHA-256 of ss_encoder.onnx: 904d66260f36d4922b891d366abddfaa48f76adafbda80126137119f6e77e8c0 (pinned in QtMeshEditor's src/ImageTo3D/ShapeGuide.cpp; the app verifies it).

I/O contract

name	type	shape	meaning
input occupancy	float32	[1,1,64,64,64]	surface voxels of the guide, 0 or 1, index (x,y,z) of the TRELLIS grid (Z up)
output latent	float32	[1,8,16,16,16]	posterior mean (no sampling)

The guide mesh is centred and scaled so its largest side spans 0.95 of the [-0.5, 0.5] cube before voxelization. In TRELLIS.2's sampler the latent is laid out [4096][8] (cell x*256 + y*16 + z, channel fastest) and mixed with noise as t0 * noise + (1 - t0) * latent, where a higher guide strength gives a lower t0.

Export

scripts/export-trellis-ss-encoder-onnx.py in the QtMeshEditor repo (offline developer tool): loads the TRELLIS checkpoint, exports with torch.onnx.export, and refuses to write a graph whose output differs from torch by more than 1e-3 (measured: 1.8e-5 on a test blockout, 8.6e-6 on random occupancy).

Licence

MIT, as the original TRELLIS code and weights (LICENSE-TRELLIS.txt).