ComfyAnonymous

ComfyAnonymous is the original creator of ComfyUI and a co-founder of Comfy Org. He built a programmable, node-based environment for assembling and altering generative-media pipelines. Users can choose how models, prompts, sampling, and image transformations work together, then preserve the process for reuse and experimentation.

From image experiments to ComfyUI

Before ComfyUI, he worked in software engineering, including web development and automation. In his account of the project’s origins, he describes discovering Stable Diffusion in October 2022 without prior experience writing PyTorch code and becoming absorbed in generating images. Early models produced relatively small images; making larger ones required generating, upscaling, and refining them in successive passes.

He wanted to experiment with those passes independently. What would happen if the second pass used a different sampler, different settings, or another model? Modifying an existing interface became more cumbersome than building his own. He started ComfyUI in January 2023 as a personal project, making the stages of generation visible and rearrangeable. His two-pass workflow examples preserve that original curiosity: one model can generate an image, an upscaler can enlarge it, and another model can refine the result.

He joined Stability AI in June 2023, where ComfyUI was used for internal model experimentation. After a year there, he left in June 2024 and started Comfy Org with collaborators. His founding statement placed free, open-source inference at the center of the organization, with image, video, and audio models as its initial priorities. It also addressed the work needed to sustain the project: improving usability, supporting new models, and making custom-node installation easier and safer.

Creative control through the generation pipeline

ComfyUI’s controls correspond to components inside the generation process. A text encoder converts a prompt into representations the diffusion model can use; a sampler iteratively constructs a result in compressed latent space; a decoder turns that representation into an image. Exposing these components separately lets users change one stage without redesigning the whole workflow. ComfyAnonymous explained this architecture in his 2025 workshop with Jedrick Kosinski.

His projects develop that control in several directions:

• Regional prompting and composition: His area-composition examples assign different prompts to different parts of an image. One combines night, evening, day, and morning in separate regions; another places subjects within a wide composition. The examples also show where control can break down: a later refinement pass without regional prompts can blend previously distinct hair colors.

• Shareable generation workflows: Saving the process is part of saving the image. His early account of ComfyUI’s design describes generated images carrying the workflow that produced them, allowing users to reload and exchange complete configurations. Someone else’s image can become the starting point for an experiment, with its graph and parameters available for inspection and modification.

• Experimental diffusion controls: His ComfyUI_experiments repository reaches below ordinary prompt settings. It includes nodes for merging model weights by block, implementing research on rescaled classifier-free guidance, and applying tonemapping to predicted noise to explore stronger guidance with less image degradation. These implementations make existing diffusion techniques available for experimentation through the node interface.

• Local inference and memory management: His explanation of ComfyUI’s backend describes estimating the memory needed for a sampling operation and unloading only enough resident model data to make room. Keeping useful data on the GPU avoids repeated loading; staying within its memory budget helps prevent severe slowdowns. This attention to execution also informs the separation of the frontend from the backend through an API.

Sharing the engine—and its maintenance

ComfyAnonymous designed ComfyUI to accommodate work beyond his own. Custom nodes let other developers extend the execution engine, while the separate frontend allows different interfaces to use it. By May 2023, he was highlighting alternative frontends and integrations alongside community extensions. The engine could support another developer’s creative tool without requiring them to adopt its node canvas.

That openness creates maintenance work as well as possibilities. In his June 2024 founding announcement, he acknowledged being overwhelmed and made better custom-node standards a priority. In December 2025, Comfy Org announced the repository’s move from his personal GitHub account to the organization, explaining that code review had long depended on him alone and that a broader team needed to share the responsibility.

ComfyUI now supports workflows across images, video, audio, and 3D, with local models and optional API-based integrations. ComfyAnonymous has also acknowledged the cost of its flexibility: node-based workflows can be difficult to learn. In the 2025 workshop, he described improving the interface and exploring a more traditional interface layer built from workflow graphs while retaining the node interface. These were directions for development, with timing dependent on a rapidly changing model ecosystem.