TRELLIS.2 Viser Visualizer¶

Use the TRELLIS.2 viser viewer to run image-to-3D inference from a browser and inspect the generated mesh interactively.

TRELLIS.2 viser visualizer

The script lives at examples/inference/trellis2_viser.py. It loads the TRELLIS.2 image-to-3D pipeline once, then lets you switch input images, tune sampler settings, run inference, and view the resulting mesh in the viser scene.

Setup¶

Install WarpConvNet with model dependencies, then add the viewer and TRELLIS runtime dependencies:

pip install "warpconvnet[models]" viser pillow safetensors huggingface_hub

The first run downloads model weights from Hugging Face:

microsoft/TRELLIS.2-4B
microsoft/TRELLIS-image-large

Use a CUDA environment with enough GPU memory for the selected pipeline mode. Pass --no_1024 if you only need the 512 mode and want to skip loading the 1024 SLAT flow.

Run¶

Launch the viewer with one or more input images:

python examples/inference/trellis2_viser.py \
    --images /path/to/image1.png /path/to/image2.webp \
    --port 8080

Or point it at a directory of images:

python examples/inference/trellis2_viser.py \
    --image_dir /path/to/images \
    --port 8080

For the lighter 512-only path:

python examples/inference/trellis2_viser.py \
    --image_dir /path/to/images \
    --no_1024

Open the printed URL, usually http://localhost:8080, in a browser.

Controls¶

Control	Description
`Image`	Selects one of the input images.
`pipeline type`	Chooses `512`, `1024`, `1024_cascade`, or `1536_cascade`.
`seed`	Sets the sampling seed.
`SS steps`	Controls sparse-structure sampler steps.
`SLAT steps`	Controls structured-latent sampler steps.
`SS guidance`	Sets sparse-structure classifier-free guidance strength.
`SLAT guidance`	Sets structured-latent classifier-free guidance strength.
`fill holes`	Toggles mesh hole filling in the shape decoder.
`Run inference`	Generates a mesh for the current image and sampler settings.
`Clear mesh`	Removes the displayed mesh from the scene.

Results are cached by image path and sampler settings inside the running process. Re-running the same configuration redraws the cached mesh instead of running the pipeline again.