ComfyUI is a node-graph engine for generative media. Users wire nodes for model loading, conditioning, sampling, and post-processing into workflows that generate images, video, 3D models, audio, and text, without writing code. It runs locally on Windows, Linux, and macOS through a desktop application, a portable build, or a manual install, and as a paid hosted service, Comfy Cloud.
Execution is incremental: only the parts of a graph with complete inputs run, and only the parts that changed since the last run are executed again. A local execution queue, partial re-execution, VRAM and RAM management, model offloading, and quantized model support target consumer hardware, and asynchronous weight streaming runs large open models on as little as 4 GB of VRAM. Supported hardware includes NVIDIA, AMD, Intel, Apple Silicon, and Ascend.
The core runs fully offline and downloads nothing unless asked. Optional partner nodes call closed models such as Nano Banana, Seedance, and Hunyuan3D through a paid API, and a flag turns them off. Workflows are saved as JSON, and generated media carries the full workflow and seeds, so dropping an output image back onto the canvas restores the graph that made it. Subgraphs, workflow templates, App Mode (which exposes a complex workflow through a simple UI), and a local API connect workflows to other applications.
| Category | Example models |
|---|---|
| Image generation | Stable Diffusion 1.5, SDXL, SD3.5, Flux.1, Flux.2, Qwen Image, HiDream, Chroma, Kandinsky 5 |
| Image editing | Flux Kontext, Qwen Image Edit, OmniGen2, HiDream E1.1 |
| Video generation | Wan 2.1 and 2.2, LTX-Video 2 and 2.3, HunyuanVideo 1.5, CogVideoX, Mochi |
| Audio generation | ACE-Step 1.5, Stable Audio 3, MiniMax Music 3 |
| 3D and vision | Hunyuan3D 2.1, SeedVR2, SUPIR, Depth Anything 3, SAM 3 |
| Text generation | Gemma 3 and 4, Qwen3, Qwen3.5, Qwen3-VL |
Custom nodes extend the engine, and ComfyUI-Manager installs, updates, and manages them. The core, the desktop application, and the frontend are developed in three linked repositories on a roughly weekly release cycle.
Features
- Node graph editor: builds image, video, audio, 3D, and text workflows visually from connected nodes
- Incremental execution: reruns only the nodes whose inputs changed since the previous run
- Low-memory inference: asynchronous weight streaming, model offloading, and quantized models run large models on consumer GPUs
- Model components: loads full checkpoints or separate diffusion models, VAEs, text encoders, LoRAs, ControlNets, adapters, and upscalers
- Editing tools: inpainting, outpainting, reference conditioning, masks and compositing, model merging, upscaling, frame interpolation, segmentation, and depth estimation
- Workflow recovery: workflows save as JSON, and generated media embeds the full graph and seeds
- Partner nodes: optional paid API nodes reach closed models such as Nano Banana, Seedance, and Hunyuan3D, and can be disabled
- App Mode and API: exposes a workflow through a simple UI or a local API for use inside other applications
- Subgraphs and templates: reusable subgraphs and a library of maintained workflow templates
- High bit depth output: saves and loads 16-bit PNG, 32-bit EXR, 10-bit AVIF, and HDR images and video
- Custom nodes: extensions add new nodes, managed through ComfyUI-Manager
- Hardware support: NVIDIA, AMD, Intel, Apple Silicon, and Ascend, plus other accelerators
