ComfyUI — Local ROCm Fork (0xra0)

Base: comfyanonymous/ComfyUI v0.38.0
Branch: local-rocm — AMD ROCm / gfx1030 (RX 6800) patches on top of upstream.

This fork contains a focused set of changes on top of ComfyUI v0.38.0 that make ComfyUI work reliably on AMD ROCm with 16 GB VRAM (gfx1030 / RX 6800), including a Pixal3D-GGUF MultiView projected-conditioning fix and an optimised launch script.

---

What's Different from Upstream

1. main.py — ROCm backend detection for comfy-aimdo

Upstream comfy-aimdo initialisation does not pass the backend implementation identifier, so the library may default to CUDA paths that fail silently on ROCm.

The patch reads torch/version.py at startup and passes "rocm" as the first positional argument to comfy_aimdo.control.init() when a ROCm PyTorch build is detected. Fallback chain for older comfy-aimdo protocols (0.4.9 / 0.4.10 / 0.5.x) is preserved.

# Before (upstream)
comfy_aimdo.control.init(simple_vram_headroom=..., nvml_pressure=...)

# After (this fork)
comfy_aimdo.control.init("rocm", simple_vram_headroom=..., nvml_pressure=...)

2. launch_best_cpu_vae.sh — Optimised ROCm launch script

A single-file launcher that sets up the environment and starts ComfyUI with settings tuned for ROCm 16 GB VRAM:

Flag	Reason
--cpu-vae --fp32-vae	Keeps the VAE on CPU to free VRAM for diffusion
--reserve-vram 2	Reserves 2 GB for OS / display driver headroom
--disable-dynamic-vram	Avoids ROCm allocator pressure spikes
--disable-async-offload	Prevents async offload race conditions on ROCm
--disable-pinned-memory	Reduces ROCm pinned-memory fragmentation
--preview-method none	Removes preview decode overhead during generation
--enable-manager	Enables the ComfyUI-Manager endpoint
PYTORCH_ALLOC_CONF=expandable_segments:False	Fixes NaN / wrong output on first Trellis2 HR flow forward on ROCm/gfx1030
OMP_NUM_THREADS=8	Caps CPU thread count for VAE and CPU-side ops

3. Pixal3D_GGUF_MultiView_Fix/ — MultiView projected-conditioning fix

A patch directory for the ComfyUI-Trellis2-GGUF custom node that fixes Pixal3D-GGUF MultiView generation on ROCm 16 GB cards.

What it fixes:

• Pixal3D-GGUF no longer falls through to TRELLIS generic spatial-blending MultiView sampling.
• Front / left / back / right views are converted to Pixal3D camera poses and fused via DinoV3ProjMultiViewFeatureExtractor.
• Fixes the proj_context=None / ProjectAttention crash.
• Uses sparse-only projected shape/texture conditioning — R=64/96 no longer allocates a full dense R³ projected feature tensor on a 16 GB ROCm card.
• NAF stays enabled; in low-VRAM mode its projected path uses tiled NAF.
• Removes deeply interior sparse voxels and aborts pathological dense samples before they consume an hour in Shape SLat.
• Corrects cascade coordinate scaling when sparse_structure_resolution != 32.

4. styles/default.csv — Default style presets

Two style presets shipped with this fork:

• ❌ Low Token — negative prompt blocking common quality issues.
• ✅ Line Art / Manga — Anime/manga positive prompt targeting Satoshi Kon, Ken Sugimori, Hiromu Arakawa aesthetics.

---

Hardware Target

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

---

Installation

Prerequisites

• Linux with ROCm 6+ drivers installed (ROCm install guide)
• Conda or a Python 3.12 venv
• At least 16 GB GPU VRAM (or adjust --reserve-vram / --disable-dynamic-vram)

1. Clone this repo

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

2. Create the Python environment

conda create -n ComfyUI python=3.12
conda activate ComfyUI

3. Install PyTorch for ROCm

Pick the ROCm wheel matching your ROCm version from PyTorch's install page.
Example for ROCm 6.2:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.2

4. Install ComfyUI dependencies

pip install -r requirements.txt

5. Configure paths

Edit launch_best_cpu_vae.sh and update:

# Line 3: change to your ComfyUI install path
cd /your/path/to/ComfyUI

# Line 7: change to your Python binary
exec /your/path/to/python3.12 main.py ...

Or set environment variables at runtime:

COMFYUI_ROOT=/your/path PYTHON=/your/python launch_best_cpu_vae.sh

6. Launch

chmod +x launch_best_cpu_vae.sh
./launch_best_cpu_vae.sh

ComfyUI will be available at http://127.0.0.1:8188 by default.
Override the port with COMFY_PORT=9000 ./launch_best_cpu_vae.sh.

---

Pixal3D-GGUF MultiView Fix (Optional)

If you use the ComfyUI-Trellis2-GGUF custom node with Pixal3D-GGUF weights, apply the MultiView fix:

Prerequisites

• ComfyUI-Trellis2-GGUF custom node installed under custom_nodes/
• Pixal3D-GGUF model weights

Install

cd Pixal3D_GGUF_MultiView_Fix
./install.sh

The installer:

1. Verifies the custom node is present.
2. Refuses to downgrade a newer installed pipeline (use FORCE=1 to override).
3. Backs up replaced files as .before-pixal3d-mv-<timestamp>.
4. Copies trellis2_image_to_3d.py, model_manager.py, and pipeline_mv.json.
5. Syntax-checks the installed files via python -m py_compile.

Verify

cd Pixal3D_GGUF_MultiView_Fix
./verify.sh

Expected workflow settings

Setting	Value
Model	Pixal3D-GGUF
Format	GGUF Q8_0
Backend	sdpa
Device	cuda (ROCm)
low_vram	true
Views	front, back, left, right
pipeline_type	512
sparse_structure_resolution	32

Expected log markers after applying the fix:

[Pixal3D-GGUF MV] Using camera-projected conditioning
[Pixal3D MV] views=['front', 'left', 'back', 'right']
[Sparse] occupied tokens after shell=...

---

Custom Nodes

Custom nodes are not included in this repository.
Install them separately via ComfyUI-Manager or by cloning them into the custom_nodes/ directory.

The .gitignore already excludes custom_nodes/ (upstream behaviour).

---

Relationship to Upstream

This branch tracks comfyanonymous/ComfyUI v0.38.0 and applies local patches on top. It is not intended as a general-purpose fork — it is a personal hardware-specific setup committed for reproducibility and sharing.

Pull requests and issues are welcome for the ROCm-specific patches but feature requests that belong in upstream ComfyUI should be directed there.

---

License

GNU General Public License v3.0 — same as upstream ComfyUI.