Tutorial Published 2026-06-18

Installing TRELLIS 2 in ComfyUI on Windows: the RTX 50-series route from driver to first mesh

TRELLIS 2 arrives as a node package plus three CUDA extensions that compile on your own machine, and an RTX 50-series card changes which toolkit and which PyTorch wheel will work. This is the install order, the environment variables, and the fixes for the errors you will actually hit.

Installing TRELLIS 2 on Windows is less an installation than a small toolchain project. The node package is ordinary Python; the cost sits in three CUDA extensions the model needs, each compiled from source on the machine that will run it. On an RTX 50-series card those builds have to target a compute capability the CUDA ecosystem learned about only recently, which is where a guide written for a 4090 stops working.

The order below wastes the least time: driver and toolkit, then the interpreter, then torch, then the extensions, then the weights. Each step has a check you can run before moving on, and the later sections map failure messages to their real causes.

Why an RTX 50-series card breaks the default TRELLIS 2 install

TRELLIS 2 is not one wheel. The node package is Python, but the computation runs through three CUDA extensions compiled on your machine at install time: cumesh for mesh operations on the decoded surface, o_voxel for the sparse voxel structure the model predicts, and flexgemm for the sparse matrix kernels underneath both.

Linux hides those three builds behind a single setup script; Windows has no equivalent. RTX 50 hardware reports compute capability 12.0, so every kernel must be emitted for sm_120. PyTorch wheels built against CUDA 12.1, 12.4 or 12.6 carry none, and a card that runs games fine stops at CUDA error: no kernel image is available for execution on the device as soon as a TRELLIS node samples. A toolkit older than CUDA 12.8 answers with nvcc fatal: Unsupported gpu architecture compute_120. Two messages, one cause: the toolchain predates the card.

What to install before ComfyUI

64-bit Windows 11 on a recent build, long path support on, and git config --system core.longpaths true: the extension sources nest past the 260-character default.
An NVIDIA driver matching your runtime, around the 570 branch for CUDA 12.8 and 580 for 13.0; nvidia-smi prints the highest CUDA version it supports.
The CUDA Toolkit itself, 12.8 or 13.0; 12.8 was the first release to accept sm_120. Confirm nvcc --version, because a stale CUDA 11.8 in PATH is common.
Visual Studio 2022 Build Tools with the C++ desktop workload, a Windows 11 SDK and ninja, plus 60-100 GB free on a path without spaces and outside OneDrive.

The install order that works

Driver, toolkit and Build Tools first, then reboot; compiling before the toolkit exists leaves a poisoned build directory behind.
Decide where ComfyUI lives. The portable build runs an interpreter inside python_embeded, a venv has its own, and every later command must use that one.
Install a cu128 or cu130 torch wheel from the PyTorch index, then confirm torch.cuda.get_device_capability() returns the pair (12, 0).
Install ComfyUI Manager and the TRELLIS 2 node package but decline the automatic requirements install, then build cumesh, o_voxel and flexgemm, clearing the build and egg-info folders between attempts.
Clear the login and license gates, download the weights, restart ComfyUI from the same launcher, and read its console rather than the browser tab.

Which CUDA and PyTorch build to pair with your card

Card	Memory (vendor spec)	Wheel and toolkit	What to expect
RTX 5090	32 GB GDDR7	torch cu128 or cu130, CUDA 12.8/13.0	Memory is no longer the constraint; the toolchain is
RTX 5080	16 GB GDDR7	The same	Works with offloading and one image per batch; the peaks are what hurt
RTX 5070 Ti	16 GB GDDR7	The same	Same capacity, fewer cores per step, so more wall-clock time
RTX 5070	12 GB GDDR7	The same	Tightest case: lower sparse-voxel resolution, offloading, texturing runs out first
Laptop RTX 50 parts	8-16 GB, shared power budget	Same wheel, another driver branch	Desktop memory figures are optimistic; throttling shows up as timeouts

Building cumesh, o_voxel and flexgemm from source

None of the three ships a Windows wheel for sm_120, and one built for an older architecture is worse than none: it imports, then falls back to a slow path or dies at the first kernel launch with a message that blames the model. Compile in the interpreter ComfyUI runs on. Read the first error in the log, not the last: a missing compiler produces a hundred cannot-open-include-file lines that vanish once cl.exe is reachable.

Environment variables that decide whether the build succeeds

CUDA_HOME pointing at the toolkit root, the v12.8 or v13.0 folder under NVIDIA GPU Computing Toolkit; the build helper reads that, not the driver.
TORCH_CUDA_ARCH_LIST set to 12.0, so nvcc emits sm_120 code instead of guessing; a guessed 8.6 module loads and then fails on the first kernel.
DISTUTILS_USE_SDK and MSSdk set to 1 after vcvars64.bat, or setuptools insists Visual C++ 14.0 is missing when it is installed.
MAX_JOBS at 4 or lower, and a build directory without spaces: three concurrent nvcc processes on a 16 GB machine reliably produces crashes that look like source bugs.

Build and runtime errors, and what each one means

nvcc fatal: Unsupported gpu architecture compute_120 means the nvcc on PATH is older than 12.8; with two CUDA installs the older one usually wins the PATH race.
no kernel image is available for execution on the device, after a successful build, means the extension targeted another architecture; inspect it with cuobjdump and rebuild.
LINK : fatal error LNK1181: cannot open input file cudart.lib means CUDA_HOME is set but the x64 library folder is not on LIB.
ModuleNotFoundError for o_voxel after a build that printed Successfully installed means you built into a different interpreter, and ImportError: DLL load failed usually means a missing Visual C++ runtime.

HuggingFace 404s, gated repositories and mirror lag

Log in from the ComfyUI interpreter, not a system shell: a read token is cached per environment, and the downloader will not see a login made elsewhere.
Accept the license in a browser with the same account; many of these checkpoints are gated, and a valid token does not pass a gate nobody accepted.
Pin a revision, because repositories get renamed and files move between subfolders, and a moved file reaches you as a plain 404.
On a mirror such as hf-mirror.com, remember that mirrors lag upstream by hours to days, and on Windows a symlinked cache download needs Developer Mode or an elevated shell.

The xformers and attention-backend mismatch

xformers wheels are pinned to one torch version and one CUDA runtime. Move to cu128 or cu130 for Blackwell and the matching build may not exist yet: the import reports that it was built for a different PyTorch and either disables itself or raises. flash-attn has the same shape of problem, and prebuilt sm_120 wheels arrive later than the torch release they target. Treat both as optional speed layers; starting ComfyUI with xformers disabled or on the PyTorch cross-attention backend costs speed, not correctness.

Verifying the install, and the case for skipping it

Verify in layers rather than by launching a workflow. torch.cuda.get_device_capability() should return the pair (12, 0); importing cumesh, o_voxel and flexgemm should raise nothing; the console should list the TRELLIS nodes at startup with no import warning above them. Only then push one image through the smallest preset and watch the console for the first error. Expect fragility across upgrades: a torch bump invalidates all four compiled components, so pin versions or keep a known-good portable copy to roll back to.

If the mesh matters more than the local stack, the same image-to-3D job can run in a browser. Open3DAI runs one local procedural generation engine and can connect to third-party engine APIs with your own keys. It does image-to-3D and text-to-3D, accepts multiview input, applies AI texturing, exports GLB, glTF, OBJ, STL, PLY and 3MF, and adds 48 format converters, 12 browser tools and a REST API; the same image can also run through the available engines side by side before you export. It will not hand you a node graph, host any third-party model, retopologize, unwrap UVs, rig, or promise watertight output. The mesh still needs cleanup, and ComfyUI stays the right choice for custom node logic, batch control or your own checkpoints.

Node packages, model repositories, driver branches and credit pricing all change; check the repository you install from and the vendor page for the engine you plan to use before committing an afternoon or a budget.

Frequently Asked Questions

Do I need the CUDA Toolkit if I only install PyTorch wheels?

A wheel-only install can run, because the PyTorch wheels carry their own runtime. You need the toolkit for the source builds: cumesh, o_voxel and flexgemm all invoke nvcc, and nvcc older than CUDA 12.8 does not accept sm_120. Install 12.8 or 13.0 and set CUDA_HOME before building.

Will TRELLIS 2 run on a 12 GB RTX 5070?

It can, with offloading and a lower sparse-voxel resolution, but the texturing stage is usually the first to run out of memory. 16 GB is the practical floor for a comfortable single-image run, and 32 GB on a 5090 removes memory from the list of problems and leaves only the toolchain.

Why does the download still return 404 after I log in?

Logging in does not accept a license gate. Open the model page in a browser with the same account, accept the terms, then retry with an explicit revision. On a mirror endpoint, remember that mirrors lag upstream, so a 404 may only mean the file has not propagated yet.

Can I skip xformers entirely?

Yes. Start ComfyUI with xformers disabled, or switch to the PyTorch cross-attention backend. You lose some speed on attention-heavy stages and nothing else, because the TRELLIS nodes run on the same torch kernels either way.

Run it yourself in Open3DAI

Send the same prompt or image to several engines, compare the meshes side by side, then export GLB, glTF, OBJ, STL, PLY or 3MF. Creating an account grants free credits and needs no card.

Start Generating Free 3D File Converter