ComfyUI: a node-graph engine for diffusion models

Products that turn a prompt into a picture hide the model behind a button. That is simple, and it also means every decision was made by someone else. ComfyUI is the infrastructure layer instead: a node graph where model loading, samplers, ControlNets and VAEs are nodes you wire together (Python, GPL-3.0, about 137.6k stars as of 2026-10-15), building pipelines for images, video, 3D and audio.

Core features

• The graph is the pipeline: no code — edges express “this model’s output feeds that node’s input”. Workflows save as JSON, can be shared and versioned, and complete workflows with seeds can be recovered from generated media.
• Very broad model coverage: image generation from Stable Diffusion 1.5 through SDXL, SD3.5, Flux.1/2 and Qwen Image; editing with Flux Kontext and Qwen Image Edit; video with Wan 2.1/2.2, LTX-Video, HunyuanVideo and CogVideoX; audio and 3D with Hunyuan3D, Depth Anything and SAM 3. The README’s list is representative — runnable templates live in the official workflow library.
• Local hardware first: the README calls itself the most optimized diffusion inference engine for local consumer hardware, using asynchronous weight streaming so that “even the biggest open source models” run “relatively quickly” on 4GB VRAM plus 8GB system RAM — the project’s own figures.
• Efficient execution: an async queue, partial graph re-execution (change one node, don’t rerun the graph), smart VRAM/RAM management, model offloading and quantized-model support.
• Fully offline capable: the core downloads nothing; --offline disables the optional paid Comfy API nodes and --disable-partner-nodes disables just the API nodes.
• Production integration: complex workflows can be wrapped into a simple UI with App Mode, and local API endpoints wire workflows into application pipelines.

Typical use cases

• E-commerce image pipelines: model shots, scene shots and multi-angle product views into workflows, run in batches with reproducible parameters.
• Creators who want control over every step: samplers, schedulers and ControlNet weights by hand instead of by luck.
• Teams embedding generation into their own products: API endpoints plus App Mode keep the pipeline in your hands while the UI stays simple for end users.

Quick start

The desktop app (Windows and macOS, from comfy.org/download) is the easiest start; manual install covers all platforms and GPU types (NVIDIA, AMD, Intel, Apple Silicon, Ascend). Once it runs, pick a template from the official workflow library and modify it — far faster than starting from an empty canvas, and the template doubles as documentation for what each node does.

Summary

For individuals and teams that want generation pipelines locked in — and for developers delivering batch image work. It is not the tool for “type text, get picture” use: the node graph has a real learning curve, and a misconfigured node does not always produce a friendly error. Python, GPL-3.0 (copyleft — check before building on it), 137,646 stars as of 2026-10-15 with pushes the same day. The release cadence is roughly fortnightly for the core repo, and the README warns that commits outside stable release tags “may be very unstable” — custom-node authors target the stable line.

If you want “topic in, video out” without building graphs, MoneyPrinterTurbo hides the generation behind a product; the related Pixelle-Video runs on ComfyUI itself — one is a packaged workflow, the other is the workflow.