ComfyUI

ComfyUI is a node-graph engine for generative media. You wire nodes into a workflow that produces images, video, 3D assets, audio, or text, and every model, parameter, and intermediate step in that graph stays under your control. It runs on Windows, Linux, and macOS across NVIDIA, AMD, Intel, Apple Silicon, and Ascend hardware, as a desktop application, a portable install, or the hosted Comfy Cloud service.

Workflows are reusable: subgraphs and templates package parts of a graph, App Mode exposes a complicated graph behind a simple form, and a local API lets other software submit workflows. Execution is asynchronous and incremental — the queue re-runs only the parts of a graph that changed, and VRAM and RAM management, model offloading, and quantized model support keep large models on modest hardware. The core runs offline and downloads nothing unless you ask; --disable-api-nodes turns off the optional paid API nodes so that stays true.

Native model support spans image generation (Stable Diffusion 1.5, SDXL, SD3.5, Flux.1 and Flux.2, Qwen Image, Z-Image, Hunyuan Image 2.1, HiDream, Chroma, and others), image editing (Flux Kontext, Qwen Image Edit, HiDream E1.1 and O1, OmniGen2), video generation (Wan 2.1 and 2.2, LTX-Video, HunyuanVideo 1.5, CogVideoX, Mochi), audio generation (ACE-Step 1.5, Stable Audio 3), 3D and vision models (Hunyuan3D 2.1, TripoSplat, SeedVR2, SUPIR, Depth Anything 3, SAM 3), and text generation (Gemma 3 and 4, Qwen3 and Qwen3-VL). Partner nodes reach closed-source models, and custom nodes extend the graph further.

Releases follow a roughly weekly cycle across three repositories — the core engine, the desktop build, and the frontend — with a new major stable core version about every two weeks and patch releases for backported fixes. Commits outside the stable tags can break custom nodes.

Features

• Visual node graph: build and reuse image, video, audio, 3D, and text workflows without writing code
• Reusable structure: subgraphs, workflow templates, App Mode, and a local API for embedding workflows in applications
• Incremental execution: asynchronous queueing and partial graph re-execution avoid recomputing unchanged branches
• Memory management: VRAM and RAM handling, model offloading, and quantized model support
• Broad model support: image, image-editing, video, audio, 3D, vision, and text model families run natively
• Component loading: full checkpoints or separate diffusion models, VAEs, text encoders, LoRAs, ControlNets, adapters, and upscalers
• Built-in editing tools: inpainting, outpainting, reference conditioning, masking and compositing, model merging, upscaling, frame interpolation, segmentation, and depth estimation
• Workflow portability: save and load workflows as JSON, or recover a full workflow and seed from generated media
• Offline operation: the core downloads nothing unless asked, and --disable-api-nodes disables the paid API nodes
• High bit-depth media: 16-bit PNG, 32-bit EXR, 10-bit AVIF, and HDR image and video formats
• Custom nodes and paths: third-party nodes extend the graph, and extramodelpaths.yaml points at additional model directories
• Deployment choices: desktop application, manual install on any supported GPU type, or the hosted cloud service