ComfyUI¶

Node-based generative AI workbench for images, video, audio, and 3D.

Overview¶

ComfyUI provides:

Node-based workflow editor - Visual graph for building generation pipelines
Multi-modal generation - Images, video, audio/music, and 3D models
Native AMD ROCm support - Official partnership with AMD as of v0.7+
Extensible - Thousands of custom nodes via ComfyUI Manager
API-driven - Every workflow is a JSON graph, fully automatable
No vendor lock-in - Open source (GPL-3.0), runs locally

Supported Modalities¶

Modality	Models	Status
Image	FLUX, SDXL, SD 1.5, SD 3.5	Mature
Video	LTX Video, HunyuanVideo, Wan 2.1/2.2	Stable
Audio/Music	ACE-Step, Stable Audio, DiffRhythm	Growing
3D	Hunyuan3D, Stable Fast 3D, Tripo	Emerging

System Requirements¶

Component	Minimum	Recommended
GPU	AMD RDNA2+ or Apple M1+	AMD RDNA 3.5 (Strix Halo) or Apple M3/M4
VRAM	8 GB	16+ GB (or unified memory APU)
RAM	16 GB	32+ GB
Python	3.10	3.12
Browser	Any modern browser	Chromium-based
Disk	20 GB (base)	100+ GB (with models)

AMD Strix Halo APU

The Ryzen AI Max+ 395 with unified memory is well suited for ComfyUI. Large models that exceed typical discrete GPU VRAM can run using shared system memory. ROCm 7.x provides native gfx1151 support -- no HSA_OVERRIDE_GFX_VERSION workaround needed.

Installation¶

Native (Recommended for AMD APU)¶

Set up a Python virtual environment with ROCm-enabled PyTorch:

# Create and activate virtual environment
python3.12 -m venv ~/comfyui-venv
source ~/comfyui-venv/bin/activate

# Install PyTorch with ROCm support
pip install torch torchvision torchaudio \
  --index-url https://repo.radeon.com/rocm/manylinux/rocm-rel-7.1/

Clone and install ComfyUI:

git clone https://github.com/comfyanonymous/ComfyUI.git ~/ComfyUI
cd ~/ComfyUI
pip install -r requirements.txt

Launch:

python3 main.py --listen 0.0.0.0

Access at http://localhost:8188

ROCm Setup

If ROCm is not yet installed, see the ROCm Installation Guide first. Ensure your user is in the video and render groups.

Docker (AMD ROCm)¶

services:
  comfyui:
    image: hardandheavy/comfyui-rocm:latest
    container_name: comfyui
    ports:
      - "8188:8188"
    devices:
      - /dev/kfd
      - /dev/dri
    group_add:
      - video
      - render
    volumes:
      - /mnt/tank/ai/comfyui/models:/app/models
      - /mnt/tank/ai/comfyui/output:/app/output
      - /mnt/tank/ai/comfyui/custom_nodes:/app/custom_nodes
    environment:
      - HIP_VISIBLE_DEVICES=0
    restart: unless-stopped

docker compose up -d

Comfy CLI¶

Quick alternative using the official CLI:

pip install comfy-cli
comfy install
comfy launch

Model Management¶

ComfyUI Manager¶

ComfyUI Manager is built into recent versions and provides:

Browse and install custom nodes from a registry
Download models directly from the UI
Resolve missing dependencies automatically
Detect and fix node conflicts
Security scanning for custom nodes

Access via the Manager button in the top menu bar.

Model Directories¶

Place models in the corresponding subdirectory under ComfyUI/models/:

Directory	Model Type	Examples
`checkpoints/`	Base models	FLUX, SDXL, SD 1.5
`loras/`	LoRA adapters	Style LoRAs, character LoRAs
`vae/`	VAE models	SDXL VAE, custom VAEs
`controlnet/`	ControlNet	Canny, depth, pose
`clip/`	CLIP text encoders	CLIP-L, CLIP-G, T5-XXL
`unet/`	UNet / diffusion models	FLUX dev/schnell UNets
`upscale_models/`	Upscalers	RealESRGAN, 4x-UltraSharp
`clip_vision/`	CLIP vision encoders	For IP-Adapter, style transfer

Shared Model Paths¶

To share models across tools, create symlinks or use ComfyUI's extra_model_paths.yaml:

# ComfyUI/extra_model_paths.yaml
shared:
  base_path: /mnt/tank/ai/models/
  checkpoints: checkpoints/
  loras: loras/
  vae: vae/

Image Generation¶

Supported Models¶

Model	Parameters	VRAM	Notes
FLUX.1 Dev	12B	16+ GB	High quality, guidance distilled
FLUX.1 Schnell	12B	16+ GB	Fast (4 steps), Apache-2.0
FLUX Kontext	12B	16+ GB	In-context editing
SDXL	3.5B	8+ GB	Mature ecosystem, many LoRAs
SD 1.5	860M	4+ GB	Lightweight, huge LoRA library
SD 3.5	2.5B-8B	8-16 GB	Latest Stability AI release

Key Capabilities¶

Text-to-image -- Type a prompt, get an image
Image-to-image -- Transform existing images with style or content changes
Inpainting -- Edit specific regions using masks
Outpainting -- Extend images beyond their borders
ControlNet -- Guide generation with edge detection, depth maps, or poses
LoRA -- Apply fine-tuned style/character adapters on top of base models
Upscaling -- Increase resolution with dedicated upscaler nodes

Basic Workflow¶

A minimal text-to-image workflow connects these nodes:

Load Checkpoint --> CLIP Text Encode (positive) --> KSampler --> VAE Decode --> Save Image
                --> CLIP Text Encode (negative) ->
                --> Empty Latent Image ->

ComfyUI ships with default workflows. Load them via Load Default or drag a workflow JSON onto the canvas.

Video Generation¶

Supported Models¶

Model	Parameters	Resolution	Notes
LTX Video	2B	Up to 768x512	Built into ComfyUI core
HunyuanVideo 1.5	13B	Up to 1280x720	High quality, text-to-video and image-to-video
Wan 2.1 / 2.2	1.3B-14B	Up to 1280x720	Multiple sizes, strong motion

Capabilities¶

Text-to-video -- Generate video clips from text descriptions
Image-to-video -- Animate a still image
Video-to-video -- Restyle or modify existing clips
Frame interpolation -- Increase FPS of generated video

Memory Requirements¶

Model	Minimum VRAM	Recommended
LTX Video 2B	8 GB	16 GB
HunyuanVideo 13B	24+ GB	48+ GB
Wan 2.1 1.3B	8 GB	16 GB
Wan 2.1 14B	24+ GB	48+ GB

APU Shared Memory Advantage

Unified memory APUs like the Ryzen AI Max+ 395 can allocate well beyond typical discrete VRAM limits. A system with 128 GB RAM can run models that would not fit on a 24 GB discrete GPU. Use --lowvram and --disable-pinned-memory flags for best results with large video models.

Audio and Music Generation¶

ACE-Step¶

ACE-Step is a 3.5B parameter music generation model capable of producing full songs:

Generate complete songs in under 10 seconds on consumer hardware
50+ language support for lyrics
Structural tags: [verse], [chorus], [bridge], [intro], [outro]
Text-to-audio, audio-to-audio, style transfer, voice cloning
Apache-2.0 license, free for commercial use

Setup:

Download ace_step_v1_3.5b.safetensors to models/checkpoints/
Install ACE-Step custom nodes via ComfyUI Manager

Basic node chain:

Load Checkpoint --> TextEncodeAceStepAudio --> KSampler --> Save Audio
                --> EmptyAceStepLatentAudio ->

Example prompt:

[verse]
Walking through the city at night
Neon lights reflecting in the rain

[chorus]
We are the dreamers, we are the light

With style tag: electronic, synthwave, 120bpm, female vocal

Stable Audio¶

Generate up to 3-minute tracks in under 2 seconds
Multi-part structure support (intros, developments, outros)
Available via Stability AI custom nodes

DiffRhythm¶

Open-source music generation alternative
Strong rhythmic consistency
Available as ComfyUI custom node pack

3D Generation¶

Supported Models¶

Model	Capability	Output	Notes
Hunyuan3D 2.0/3.0	Text-to-3D, image-to-3D	UV-mapped meshes	Tencent, high fidelity
Stable Fast 3D	Single image to mesh	Textured mesh	~500ms generation
Tripo v3.0	Text/image-to-3D	PBR materials	High-fidelity meshes

Output Formats¶

UV-mapped meshes (OBJ, GLB)
PBR material textures (albedo, normal, roughness)
Ready for import into Blender, Unity, Unreal Engine

Install 3D generation nodes via ComfyUI Manager. Most require the trimesh and pytorch3d Python packages.

Configuration¶

Launch Parameters¶

Flag	Description	When to Use
`--listen 0.0.0.0`	Listen on all interfaces	Network access
`--port 8188`	Set port number	Change default port
`--lowvram`	Aggressive memory optimization	Large models, limited VRAM
`--disable-pinned-memory`	Disable HIP pinned memory	APU unified memory
`--cpu`	Force CPU inference	No GPU available
`--force-fp16`	Force FP16 precision	Save memory
`--preview-method auto`	Enable live previews	See generation progress

Recommended for AMD APU:

python3 main.py --listen 0.0.0.0 --lowvram --disable-pinned-memory --force-fp16

Environment Variables¶

For AMD ROCm:

export HIP_VISIBLE_DEVICES=0
export PYTORCH_HIP_ALLOC_CONF=expandable_segments:True
export HSA_ENABLE_SDMA=0

Custom Nodes¶

Install via ComfyUI Manager (recommended) or manually:

cd ~/ComfyUI/custom_nodes
git clone https://github.com/author/node-pack.git
pip install -r node-pack/requirements.txt

Restart ComfyUI after installing nodes.

Popular custom node packs:

Node Pack	Purpose
rgthree-comfy	Quality-of-life nodes (reroute, mute, bookmark)
ComfyUI-Impact-Pack	Detection, segmentation, and batch processing
ComfyUI-KJNodes	Utility nodes for advanced workflows
ComfyUI-VideoHelperSuite	Video loading, saving, and frame manipulation
ComfyUI-AnimateDiff	Animation and video generation helpers

API Usage¶

Every ComfyUI workflow is a JSON graph that can be queued via the REST API.

Queue a Prompt¶

import requests
import json
import uuid

def queue_prompt(workflow: dict, server: str = "http://localhost:8188") -> str:
    """Queue a ComfyUI workflow and return the prompt ID."""
    response = requests.post(
        f"{server}/prompt",
        json={
            "prompt": workflow,
            "client_id": str(uuid.uuid4()),
        },
    )
    return response.json()["prompt_id"]

Retrieve Output¶

import time

def get_output(prompt_id: str, server: str = "http://localhost:8188") -> bytes:
    """Poll for completion and return the first generated image."""
    while True:
        response = requests.get(f"{server}/history/{prompt_id}")
        history = response.json()

        if prompt_id in history:
            outputs = history[prompt_id]["outputs"]
            for node_id, output in outputs.items():
                if "images" in output:
                    filename = output["images"][0]["filename"]
                    img = requests.get(
                        f"{server}/view",
                        params={"filename": filename},
                    )
                    return img.content
        time.sleep(0.5)

WebSocket Connection¶

For real-time progress updates, connect via WebSocket:

import websocket

ws = websocket.create_connection("ws://localhost:8188/ws?clientId=my-client")
while True:
    msg = json.loads(ws.recv())
    if msg["type"] == "progress":
        print(f"Step {msg['data']['value']}/{msg['data']['max']}")
    elif msg["type"] == "executed":
        print("Done")
        break

Performance Tips¶

Use --lowvram on APU systems to aggressively offload between nodes
Use --disable-pinned-memory on unified memory systems
Use --force-fp16 to halve memory usage with minimal quality loss
Keep models loaded -- ComfyUI caches loaded models in VRAM between runs
Reduce resolution first -- Generate at lower res, then upscale with a dedicated node
Batch wisely -- Small batches avoid memory spikes on memory-constrained systems
Monitor GPU usage -- Run rocm-smi during inference to verify the iGPU is active

Troubleshooting¶

ROCm Not Detected¶

# Verify driver is loaded
lsmod | grep amdgpu

# Check device nodes exist
ls -la /dev/kfd /dev/dri/render*

# Verify group membership
groups | grep -E "video|render"

# Test PyTorch sees the GPU
python3 -c "import torch; print(torch.cuda.is_available())"

See ROCm Installation for full setup instructions.

Out of Memory¶

Add --lowvram and --disable-pinned-memory flags
Use --force-fp16 to reduce memory usage
Reduce image resolution (start with 512x512, upscale later)
Close other GPU-using applications
For video models, reduce frame count or resolution

Slow Generation¶

Verify GPU is being used: rocm-smi should show activity during inference
Check that PyTorch is using the correct device (not falling back to CPU)
Reduce inference steps (20-30 is often sufficient for FLUX Schnell)
Enable --preview-method auto only when needed (adds overhead)

Custom Node Conflicts¶

Use ComfyUI Manager to check compatibility between installed nodes
Disable recently installed nodes if ComfyUI fails to start
Check the terminal/console output for Python import errors
Remove conflicting nodes from custom_nodes/ and restart

Black or Corrupt Output¶

Verify the correct VAE is connected in the workflow
Check model format matches expected type (e.g., SDXL checkpoint with SDXL workflow)
Ensure the model file downloaded completely (check file size)
Try a different sampler or scheduler

ComfyUI¶

Overview¶

Supported Modalities¶

System Requirements¶

Installation¶

Native (Recommended for AMD APU)¶

Docker (AMD ROCm)¶

Comfy CLI¶

Model Management¶

ComfyUI Manager¶

Model Directories¶

Shared Model Paths¶

Image Generation¶

Supported Models¶

Key Capabilities¶

Basic Workflow¶

Video Generation¶

Supported Models¶

Capabilities¶

Memory Requirements¶

Audio and Music Generation¶

ACE-Step¶

Stable Audio¶

DiffRhythm¶

3D Generation¶

Supported Models¶

Output Formats¶

Configuration¶

Launch Parameters¶

Environment Variables¶

Custom Nodes¶

API Usage¶

Queue a Prompt¶

Retrieve Output¶

WebSocket Connection¶

Performance Tips¶

Troubleshooting¶

ROCm Not Detected¶

Out of Memory¶

Slow Generation¶

Custom Node Conflicts¶

Black or Corrupt Output¶

See Also¶