ComfyUI

ComfyUI is a node graph interface for building and running generative AI pipelines: image, video, 3D, and audio models connected as a visual workflow rather than a single fixed UI. Instead of a chat box or a form with sliders, you wire together nodes for loading models, encoding prompts, sampling, and post-processing, which lets you build anything from a simple text-to-image workflow to a complex multi-stage pipeline. It's built for people who want fine control over every parameter of a Stable Diffusion-style pipeline: hobbyists experimenting with local generation, and technical users building production pipelines around AI-generated media. It runs on Windows, Linux, and macOS, locally or through a paid cloud offering, and works fully offline unless you deliberately enable paid API nodes for closed-source models.

Key features

• Node graph workflow editor: chain models, samplers, and post-processing steps visually instead of coding or clicking through fixed forms.
• Broad model support: works natively with recent open image models (SDXL, SD3/3.5, Flux, Flux 2, Qwen Image, HiDream, and more), video models (Wan 2.1/2.2, Hunyuan Video, LTX-Video, Mochi, Stable Video Diffusion), audio models (Stable Audio, ACE Step), and 3D generation (Hunyuan3D 2.0).
• App Mode: exposes a complex workflow through a simplified UI so it can be handed to non-technical users.
• Smart memory management: automatically offloads parts of large models so they run on GPUs with as little as 1GB of VRAM, and can fall back to CPU-only execution.
• Optional API nodes: reach paid closed-source models (like Nano Banana and Seedance) from inside a workflow, disabled entirely with a --disable-api-nodes flag if you want to stay fully local.
• Workflow persistence: save and load full workflows, including seeds, as JSON files or embedded directly in generated PNG, WebP, and FLAC files.
• Extra techniques: supports LoRA (regular, LoCon, LoHa), textual inversion embeddings, ControlNet/T2I-Adapter, inpainting, area composition, hires fix, model merging, and upscaling models.

Ideal use cases

ComfyUI suits anyone who wants to control every step of an image, video, or audio generation pipeline rather than accept a single opinionated interface: artists building repeatable custom workflows, developers wiring generation into a production pipeline through its API endpoints, and researchers experimenting with new model combinations. The App Mode feature also makes it workable as the backend for a simplified tool aimed at less technical users, once someone has built the underlying workflow.

It's a poor fit if you just want to type a prompt and get an image with no interest in the mechanics. The node graph has a learning curve, and building non-trivial workflows takes real time investment compared to a single-prompt interface. It's also not the right choice if you have no GPU and need fast results; CPU-only mode is supported but explicitly noted as slow, and the more capable video and 3D models need meaningful VRAM even with offloading.

Installation

The easiest path is the desktop application, available for Windows and macOS. Windows users can also grab a portable, no-install package from the releases page, available in NVIDIA, AMD, and Intel GPU variants.

To install with comfy-cli:

pip install comfy-cli
comfy install

For a manual install on Windows or Linux, clone the repo, then install a PyTorch build matching your hardware. For NVIDIA:

pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130

AMD (Linux, ROCm) and Intel (XPU) each have their own PyTorch index URLs documented in the README. After that, install the remaining dependencies:

pip install -r requirements.txt

Then run it with:

python main.py

Apple Silicon (M1 through M4) works with the PyTorch nightly build following Apple's Metal-accelerated PyTorch guide, then the same manual install steps. Ascend NPUs, Cambricon MLUs, and Iluvatar Corex GPUs are also supported through their respective vendor toolkits, documented separately in the README.