Skip to content

About

ComfyUI-Trellis2 fork with AMD ROCm/gfx1030 patches: DINO/NAF stability, natten shim, sparse VAE fixes for 16 GB VRAM

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

🌀 ComfyUI-Trellis2 — Local ROCm Fork (0xra0)

Base: visualbruno/ComfyUI-Trellis2
Branch: local-rocm — AMD ROCm / gfx1030 (RX 6800) patches on top of upstream.

ComfyUI custom node wrapping microsoft/TRELLIS.2 (non-GGUF, full-precision path).
This fork adds ROCm-specific DINO/NAF stability fixes, a natten shim for platforms where natten is unavailable, and sparse VAE fixes targeting 16 GB VRAM on AMD gfx1030.


What's Different from Upstream

Commit: local-rocm-20261015 — DINO/NAF stability, natten shim, sparse VAE

File Change
nodes.py ROCm-safe node registrations; MultiView node fixes
trellis2/models/sc_vaes/sparse_unet_vae.py Sparse UNet VAE stability on ROCm (dtype/device guards)
trellis2/trainers/flow_matching/mixins/_natten_shim.py (new) natten compatibility shim — provides a pure-PyTorch fallback when the natten package is not installed or not available for the current platform
trellis2/trainers/flow_matching/mixins/image_conditioned_proj.py DINO/NAF projected conditioning stability; MultiView camera pose routing for Pixal3D

Unstaged additions (also active in the working tree):

  • nodes.py — additional MultiView node refinements
  • trellis2/trainers/flow_matching/mixins/image_conditioned_proj.py — further projection conditioning tweaks

Hardware Target

Component Spec
GPU AMD Radeon RX 6800 (gfx1030)
VRAM 16 GB
ROCm 6.x / 7.x
Python 3.12
PyTorch ROCm build

Installation

Prerequisites

  • ComfyUI installed and running (see 0xra0/ComfyUI for the matching ROCm-tuned base)
  • ROCm 6+ drivers
  • Python 3.12 environment used for ComfyUI

1. Clone into custom_nodes

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/0xra0/ComfyUI-Trellis2.git
cd ComfyUI-Trellis2
git checkout local-rocm

2. Install requirements

pip install -r requirements.txt

Note

natten (Neighborhood Attention) is optional. If not available for your platform/ROCm version, the included _natten_shim.py provides a pure-PyTorch fallback automatically — no manual configuration needed.

3. Download models

DINOv3 (required):

git clone https://huggingface.co/facebook/dinov3-vitl16-pretrain-lvd1689m \
  /path/to/ComfyUI/models/facebook/dinov3-vitl16-pretrain-lvd1689m

Trellis2 weights — follow the upstream visualbruno/ComfyUI-Trellis2 model instructions.


Usage

This node provides the full-precision (non-GGUF) Trellis2 pipeline. It is compatible with the same workflows as upstream, with the added ROCm stability layer.

DINO/NAF Stability

The upstream DINO feature extractor and NAF (Noise-Aware Fusion) path can produce NaN or crash on ROCm due to attention backend differences. This fork guards those paths with ROCm-safe alternatives while keeping the same output contract.

natten Shim

When natten is not installed, _natten_shim.py transparently replaces Neighborhood Attention calls with an equivalent SDPA-based implementation. This allows Trellis2 to run on ROCm without building the natten CUDA/ROCm extension.

MultiView / Pixal3D

The image_conditioned_proj.py changes align the projection conditioning with the camera-pose routing developed in the GGUF fork, so both nodes can be used in the same ComfyUI workflow.


Rebasing onto New Upstream

git remote add upstream https://github.com/visualbruno/ComfyUI-Trellis2.git
git fetch upstream
git rebase upstream/main

Relationship to Upstream

This is a personal hardware-specific fork of visualbruno/ComfyUI-Trellis2.
ROCm-specific issues belong here. Everything else belongs upstream.


License

Same license as upstream — see the upstream repository for the applicable license.

About

ComfyUI-Trellis2 fork with AMD ROCm/gfx1030 patches: DINO/NAF stability, natten shim, sparse VAE fixes for 16 GB VRAM

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages