comfy-org/comfyui

ComfyUI is an interface for running and combining AI generation models that produce images, video, audio, and 3D content. Unlike apps that hide settings behind simple buttons, it shows the entire generation process as a visual flowchart of nodes connected with lines. Each node is a small step (load a model, set a prompt, sample an image, save the file), and you wire them together to build the full workflow. This makes it appealing to people who want full control over every model and parameter. Under the hood it is written in Python with PyTorch, and it natively supports a long list of open-source diffusion models for images, image editing, video, audio, and 3D, plus optional API nodes that connect to paid closed-source models. It works on Windows, Linux, and macOS, and supports GPUs from NVIDIA, AMD, Intel, Apple Silicon, and Ascend, with smart memory management that can run large models on as little as 1GB of VRAM by offloading. There is a CPU-only mode for machines without a GPU. It can load checkpoints, LoRAs, embeddings, ControlNets, and upscalers, and supports advanced techniques like inpainting, area composition, and model merging. Workflows can be saved as JSON files or even reloaded from generated PNG, WebP, and FLAC files. You would use ComfyUI to design custom AI generation pipelines: complex multi-step image creations, video pipelines, batch jobs, or production setups exposed through its API. It runs locally with a desktop application, a portable Windows package, or a manual install, and there is also a paid Comfy Cloud option. The full README is longer than what was provided.