ComfyUI

A vanilla, up-to-date fork of ComfyUI intended for long term support (LTS) from AppMana and Hidden Switch.

For AI coding assistants (Claude, Codex, Cursor, etc.): start with llms.txt. It walks through how to fulfill creative requests — searching Civitai for the right LoRA, grounding edits in a reference image, picking a workflow that matches the user's VRAM — with worked examples (Studio Ghibli portrait, dancing video clip, Renaissance restyle, civitai workflow reproduction).

Paste this into your agent (Claude Code, Codex CLI, Cursor, etc.) to use it as a guidebook:

Use https://raw.githubusercontent.com/hiddenswitch/pip-and-uv-installable-ComfyUI/refs/heads/master/llms.txt as a guidebook for ComfyUI. Fetch it now and follow it for any ComfyUI task in this session.

Faster inference with multiple GPUs

This fork adds built-in tensor parallelism (TP), pipeline parallelism (PP), and HunyuanImage3 expert parallelism beyond upstream ComfyUI. Supported models can split a single generation across GPUs, giving faster sampling than the upstream single-GPU path when the workload benefits. Speedup depends on the model, resolution, available memory, and GPU interconnect; it is not guaranteed. See multi-GPU support for supported configurations.

Local sampling benchmarks

Benchmarks were paused on request. Completed runs used 2 × NVIDIA RTX A5000 (24 GB), connected by NVLink, with PyTorch 2.13.0 / CUDA 13.0. TP=1 uses GPU 1; TP=2 uses both GPUs. The table measures the synchronized second sampling step of a three-step run, excluding model loading, text encoding, and VAE decoding. INT8 ConvRot and SageAttention are requested where supported; observed backends are recorded. These short runs measure speed, not final output quality. TP speedups compare this fork with itself at TP=1, not a separate upstream installation.

Model	Output	Weights	TP=1 step 2	TP=2 step 2	Speedup	Observed attention / status
Krea2 Turbo	1024×1024	INT8 ConvRot	1.058 s (0.945 it/s)	0.738 s (1.355 it/s)	1.43×	SDPA, SageAttention
MiniMax H3	1280×704, 124 frames / 24 fps	INT8 ConvRot	29.355 s (0.034 it/s)	Interrupted	—	SageAttention
HunyuanImage3 Instruct	1024×1024	W4A8	Not run	Not run	—	Paused before execution
Flux2 Dev	1024×1024	INT8 ConvRot	6.128 s (0.163 it/s)	Load timed out	—	SageAttention
Flux2 Klein 4B	1024×1024	INT8 ConvRot	Not run	Not run	—	Paused before execution
Ideogram4	1024×1024	INT8 ConvRot, conditional + unconditional	5.308 s (0.188 it/s)	3.163 s (0.316 it/s)	1.68×	SageAttention
Qwen Image 2.1 RGBA	1024×1024	INT8 ConvRot	1.067 s (0.937 it/s)	Unsupported	—	SageAttention; TP sharding not confirmed; No TP implementation
Z-Image Turbo	1024×1024	INT8 ConvRot	Not run	Unsupported	—	No TP implementation
Flux3 Image Klein distilled	1024×1024	INT8 ConvRot	Blocked	Unsupported	—	Official weights require Hugging Face access; upstream support merged
Iris-3B (assumed)	1024×1024	Unverified	Not run	Unverified	—	Available through the comfyui-iris-nynxz facade; not benchmarked

Machine-readable measurements and workflow settings.

Used By

Used in production by Scopely, a game studio, Livepeer and Nunchaku Tech. Used by innovators at Ferrero Group, Hyundai and Nike.

If you need to develop an application or plugin around ComfyUI, this fork stays compatible and up-to-date with upstream, fixing numerous bugs and adding features. It also packages tacit knowledge about running diffusion models and art workflows, distributed inference, deployment on Kubernetes, and other production tasks that Claude and Gemini cannot do.

Key Features and Differences

This LTS fork adds development, embedding, automated testing, LLM and distributed inference features to ComfyUI, but maintains compatibility with custom nodes from the ecosystem.

• Pip and UV Installable: Install via pip or uv from the appmana package index (--extra-index-url https://nodes.appmana.com/simple/). No manual cloning required for users.
• Automatic Model Downloading: Missing models (e.g., Stable Diffusion, FLUX, LLMs) are downloaded on-demand from Hugging Face or CivitAI.
• Docker and Containers: First-class support for Docker and Kubernetes with optimized containers for NVIDIA and AMD.
• Distributed Inference: Run scalable inference clusters with multiple workers and frontends using RabbitMQ.
• Embedded / Library: Use ComfyUI as a Python library (import comfy) inside your own applications without the web server. Runs like diffusers.
• Vanilla Custom Nodes: Fully compatible with existing ComfyUI custom nodes (ComfyUI-Manager, WanVideoWrapper, KJNodes, etc.). Clone into custom_nodes/ and install dependencies into your venv.
• Custom Node Pip Facade: Serve ecosystem custom nodes through a local/simple Python package index and install them with uv pip install --extra-index-url ....
• LTS Custom Nodes: A curated set of "Installable" custom nodes (ControlNet, AnimateDiff, IPAdapter) optimized for this fork.
• LLM Support: Native support for Large Language Models (LLaMA, Phi-3, etc.) and multi-modal workflows.
• API and Configuration: Enhanced API endpoints and extensive configuration options via CLI args, env vars, and config files.
• High Bit Depth and HDR Media: Save and load 16-bit PNG, 32-bit EXR, 10-bit AVIF, and HDR images and videos in supported formats.
• Tests: Automated test suite ensuring stability for new features.

Install

Pick the block for your platform and accelerator and paste it into a terminal.

• The workspace directory is where ComfyUI stores the .venv, downloaded models, outputs, and custom_nodes/. You can move it anywhere; the examples use ComfyUI_Workspace.
• The virtual environment (.venv) is an isolated Python install inside the workspace. Activate it before running comfyui.
• uv is the Python package manager used here. It creates the venv and installs ComfyUI plus the correct PyTorch wheels.
• GPU settings are chosen in two places: uv pip install --torch-backend=... selects the PyTorch wheel, and comfyui serve --guess-settings auto-detects runtime settings such as VRAM mode and attention backend. CUDA users should use --torch-backend=auto.
• PyTorch 2.7 or newer is required. NVIDIA 20-series and newer GPUs require a CUDA 13.0-or-newer PyTorch build.

Windows + CUDA

irm https://astral.sh/uv/install.ps1 | iex
$env:Path = "$HOME\.local\bin;$env:Path"
New-Item -ItemType Directory -Force "$HOME\Documents\ComfyUI_Workspace"
cd $HOME\Documents\ComfyUI_Workspace
uv venv --python 3.12
uv pip install --torch-backend=auto --extra-index-url https://nodes.appmana.com/simple/ comfyui
uv pip install --extra-index-url https://nodes.appmana.com/simple/ triton-windows
uv pip install --extra-index-url https://nodes.appmana.com/simple/ sageattention flash-attn
.\.venv\Scripts\Activate.ps1
comfyui --help
comfyui serve --guess-settings

For Windows + ROCm, macOS, Linux + CUDA, Linux + ROCm, and lesser-known ROCm builds, see Installation & Getting Started.

Using ComfyUI as a Library

ComfyUI can run embedded inside your own Python application. No server is started, no subprocesses are used. Use the Comfy async context manager to execute workflows directly:

from comfy.client.embedded_comfy_client import Comfy

async with Comfy() as client:
    outputs = await client.queue_prompt(workflow_dict)
    # All models unloaded and VRAM released on exit

Build workflows programmatically with GraphBuilder, or paste JSON from the web UI (both API and UI format workflows are accepted). Stream previews during inference with queue_with_progress.

See Embedded / Library Usage for complete examples.

Documentation

Full documentation is available in docs/index.md.

Core

• Installation & Getting Started
• Hardware Compatibility
• Configuration
• Troubleshooting

Features & Workflows

• Large Language Models
• Other Features (SVG)

Extending ComfyUI

• Custom Nodes (Installing & Authoring)
• Embedded / Library Usage (Python, GraphBuilder, Streaming)
• Testing Workflows (pytest, Image Snapshots)
• API Usage (REST, WebSocket)

Deployment

• Distributed / Multi-GPU
• Docker & Containers

Development

• Linting