The official Microsoft TRELLIS.2 repository currently requires Linux and an NVIDIA GPU with at least 24 GB of memory for local inference. Microsoft reports verification on A100 and H100 GPUs and recommends CUDA Toolkit 12.4. Consumer-GPU timings and lower-memory configurations are not official requirements, so this guide does not invent them.
Verified local requirements
| Component | Official requirement or guidance |
|---|---|
| Operating system | Tested only on Linux |
| GPU vendor | NVIDIA |
| GPU memory | At least 24 GB |
| Verified hardware | NVIDIA A100 and H100 |
| CUDA Toolkit | Required; 12.4 recommended |
| Python | 3.8 or newer |
| Environment manager | Conda recommended |
| Pretrained model | microsoft/TRELLIS.2-4B |
Source: Microsoft TRELLIS.2 README.
What “24 GB required” does and does not mean
The upstream statement is a support boundary, not a benchmark table for every NVIDIA GPU:
- It does not promise that every 24 GB card will build every CUDA extension successfully.
- It does not document native Windows or macOS operation.
- It does not validate 6 GB, 8 GB, 12 GB, or 16 GB consumer cards.
- It does not guarantee the same runtime at every output resolution.
Community workarounds may exist, but they should be evaluated separately from official support.
Official generation-time reference
Microsoft publishes the following measurements on an NVIDIA H100:
| Output resolution | Reported total time | Shape + material |
|---|---|---|
| 512³ | about 3 seconds | 2s + 1s |
| 1024³ | about 17 seconds | 10s + 7s |
| 1536³ | about 60 seconds | 35s + 25s |
These are H100 measurements, not promises for a workstation, consumer GPU, or hosted provider. Queue time, image preprocessing, export, network transfer, and provider load are outside that table.
Source: official project performance section.
Software considerations
The installation compiles several performance-sensitive packages, including O-Voxel, CuMesh, FlexGEMM, nvdiffrast, and nvdiffrec. Local setup therefore depends on more than raw VRAM:
- NVIDIA driver compatibility
- CUDA Toolkit and
CUDA_HOME - matching PyTorch CUDA build
- compiler toolchain and Linux packages
- FlashAttention support, or the documented
xformersalternative
Use the repository's setup.sh instead of combining commands from the older TRELLIS project.
Storage and RAM
The upstream README does not currently publish a formal minimum system-RAM or free-disk number. Leave enough space for the Conda environment, compiled extensions, model weights, caches, and generated assets, but do not treat third-party 16/32/64 GB tables as official specifications.
Web generation requirements
For the trellis2.com web generator, the local requirements are different:
- a current browser
- a supported source image
- network access for upload and result delivery
- sufficient credits for the selected model and options
No local CUDA installation is required because generation runs remotely.
Use the browser-based generator
Choose local or web
Choose local installation when you need source-level control, can maintain the Linux/CUDA environment, and have supported hardware. Choose the web workflow when you want upload, generation, GLB preview, and download without maintaining a GPU stack.