ComfyUI (Comfy Org)

ComfyUI is an open-source interface and backend for running image and video generation models as node graphs. Instead of filling in a form, a user builds a workflow on a canvas: one node loads a model, another encodes the prompt, others sample, decode, upscale and save, and every step stays visible and adjustable. It was started in January 2023 by the developer known on GitHub as comfyanonymous, who is still its most prolific contributor, and is now maintained by the company Comfy Org. It is licensed under GPL-3.0. On October 24, 2026 its repository had 142,520 GitHub stars and 16,803 forks, about twice as many stars as Remotion. Its latest GitHub release was v0.42.1 (October 19, 2026), with tags up to v0.42.5 by October 23.

How It Works

Each node does one job and passes its output to the next. Because the graph is explicit, the same workflow can be shared, rerun and tweaked: swap a model, add a LoRA, insert a control image or an upscaler, and the rest stays the same. Workflows are saved as JSON and can be exported in an API format that other programs submit to a running ComfyUI server. Custom nodes extend what it can do, and Comfy Org says contributors have built more than 60,000 of them.

ComfyUI runs locally on the user's own GPU, and new open models often arrive with ComfyUI workflows: Lightricks, for example, had ComfyUI templates ready when it launched LTX-2.5. Image models such as Stable Diffusion and FLUX and video models such as Wan, HunyuanVideo and Lightricks' LTX are all run this way, and community fine-tunes and LoRAs are commonly distributed for it (see LTX vs Wan and open-weight models).

Partner Nodes

Not every model is open. Partner Nodes (also called API Nodes) let a graph call closed, hosted models, including Google's Veo and Kling, from inside ComfyUI. They run on the provider's servers, not the user's GPU, need no separate provider API keys, and are paid for with prepaid Comfy credits through a logged-in Comfy account, so they do not work offline. A single workflow can therefore mix a local open model with a hosted one.

That makes ComfyUI a common backend for other tools. The open-source agent pipeline OpenMontage, for example, sends workflows to a local ComfyUI server, including bundled Wan 2.2 workflows, and can also call hosted video models through Partner Nodes.

Comfy Org

Comfy Org builds products around the open-source core: Comfy Desktop for running ComfyUI locally, Comfy Cloud for running it hosted, a developer platform and API for deploying workflows as production APIs, an enterprise offering, and Comfy Agent, which builds workflows from a text description inside ComfyUI. On April 24, 2026 it announced a $30 million round at a $500 million valuation, led by Craft with Pace Capital, Chemistry and TruArrow, bringing its total funding to $47 million. It reported more than 4 million users and more than 150,000 downloads a day, and said the money would go toward cloud infrastructure, team collaboration, the local app and same-day support for new models.

How It Compares

Hosted creator apps such as Higgsfield, Runway and Midjourney trade control for convenience: a prompt box and presets rather than a graph. ComfyUI offers the opposite trade. It gives full control over every step, runs locally with open models at no per-image cost beyond the hardware, and supports new open models quickly, but its graphs take time to learn and a capable GPU helps. Its GPL-3.0 license covers ComfyUI itself; the models it runs carry their own licenses, which vary. For the techniques underneath, see diffusion models, text-to-image and text-to-video.