ComfyUI

ComfyUI is an open-source, node-based workflow platform for generative AI, supported by Comfy Org. Creators connect models and processing stages into pipelines for images, video, 3D, and audio, running them locally or through Comfy Cloud. On April 24, 2026, Comfy announced $30 million in financing at a $500 million valuation, alongside plans to expand cloud infrastructure, collaborative workflows, and the reliability of its community node ecosystem.[6][5][11]

The platform began with a narrower problem: experimenting with successive passes through diffusion models. Its creator’s earlier work explains why ComfyUI exposes the individual stages of generation as a graph.[1]

In the Latent Space interview published January 4, 2025, comfyanonymous recalled discovering Stable Diffusion in October 2022. He began with AUTOMATIC1111 and modified its high-resolution workflow: generate a low-resolution image, upscale it, then refine it through another diffusion pass. After varying the second pass’s sampler and settings, he wanted to try a different model. He said building his own interface was easier than implementing that experiment in the existing one. He began writing ComfyUI on January 1, 2023, and released its first GitHub version on January 16.[1]

One early experiment was area conditioning: applying different prompts to different image regions, such as placing a mountain in one area and a fox in another. Comfy described combining regional and whole-image predictions by averaging them at each diffusion step. He recalled that a Reddit post about the technique brought early attention, followed by a video from Olivio in March 2023.[1]

Comfy said Stability hired him in June 2023 because its SDXL work involved chaining a base model and a refiner—an experiment ComfyUI already accommodated. He later credited SDXL support and the ability to run it on ordinary local GPUs with increasing adoption. The April 2026 financing announcement identifies Yannik Marek as ComfyUI’s creator and a cofounder.[1][5]

ComfyUI separates generation into connected components: text input, CLIP encoding, checkpoint loading, sampling, and VAE decoding. This lets users chain stages instead of manually switching tools, and restart execution partway through a workflow. The creator described his original design choice as “let me try to make a powerful interface that's not easy to use.”[1]

Local execution shaped the backend. In the 2025 interview, Comfy explained that memory management estimated the space needed for sampling and freed GPU memory while trying to retain useful loaded models, reducing repeated transfers. He also advocated separating frontend and backend through a defined API. The repository dates the transition to a separately maintained TypeScript/Vue frontend to August 15, 2024.[1][3]

Other developers used ComfyUI as an application backend, including a Krita plugin highlighted by Comfy. Custom nodes extended its reach, while the Comfy Registry provided a library for discovering them. Comfy said the registry was intended to reduce the manual work of the ComfyUI Manager maintainer, who had searched GitHub for new nodes. Latent Space’s community resources, assembled with Major Tom, also highlighted Alessandro Perili’s AP Workflows and Kijai’s integrations for video and animation models.[1]

At the time of the interview published in January 2025, Comfy described himself as responsible for the core, with hiring underway and team attention concentrated on the previously neglected frontend. He outlined easier installation and possible cloud inference and enterprise offerings as ways to commercialize the platform while continuing open-source development.[1]

Comfy Cloud exited beta on March 4, 2026. Its billing model charged for active GPU time while a workflow ran, leaving graph editing and parameter changes unmetered. Comfy described the service’s hardware at that launch as NVIDIA Blackwell RTX 6000 Pro GPUs with 96GB of VRAM and 180GB of system RAM. Production workflow API deployment and parallel execution remained roadmap items in the announcement.[10]

On March 10, 2026, Comfy introduced App Mode, App Builder, and URL sharing on both Local and Cloud. App Mode changes the interface while preserving the workflow, backend, instance, and queue. App Builder selects which node inputs and outputs users see: a complex generation graph might expose only a prompt, seed, and input image, with the remaining settings fixed inside the workflow. Cloud recipients could open shared apps without installing software.[9]

The same launch introduced a ComfyHub preview for sharing finished apps and workflows, distinct from the Node Registry’s role in publishing custom nodes. In the accompanying community announcement, Yoland from Comfy Org described the features as beta and said the hub initially accepted a selected group of publishers to avoid moderation needs.[9][8]

Comfy announced a $1 million “Open AI” grant program for open-source AI, alongside its first sponsored model, Anima. The open-weights model was created through collaboration between CircleStone Labs and Comfy Org. Beyond funding, the program offered real-world evaluation, production testing, and promotion. Its announced terms let recipients retain control of their model and license, provided it remained open, and enroll in Cloud revenue sharing.[11]

On September 16, 2025, Comfy announced $17 million in funding, naming Pace Capital, Chemistry, and Abstract Ventures among its investors. It identified node stability, a refined interface, cloud access for users with limited local compute, and new-model support as priorities. The team also credited Tim and Naomi at Essence VC with helping shape its strategy.[4]

The April 24, 2026 financing announcement named Craft as lead investor, with Pace Capital, Chemistry, and TruArrow participating. Comfy reported 4 million users and more than 60,000 contributor-built nodes; these were ecosystem counts, without a stated active-user definition. A company community announcement separately reported that Comfy Cloud and Partner Nodes together had crossed $10 million in annualized bookings in eight months.[5][7]

Comfy’s April announcement described Black Math connecting motion, texture, and generative tools into modular creative pipelines, and Silverside AI using ComfyUI in developing SVEDKA’s 2026 Super Bowl commercial. These examples placed reusable workflows within studio and advertising production, beyond individual image generation.[5]

The financing plan paired cloud expansion and team collaboration with continued investment in local updates, node reliability, and model integrations. Comfy renewed its commitment that users would retain the ability to run the platform on their own machines with access to their work: “You will always be able to run ComfyUI on your own machine, on your own terms, with full access to everything you’ve built.”[5]