comfyanonymous

comfyanonymous, also known as Comfy, created ComfyUI to experiment with diffusion models and chain them together. His defining engineering choices were a node graph that exposed the generation process and a backend designed to run well locally. In his Latent Space interview, published January 4, 2025, he described the interface goal as “a powerful interface that's not easy to use.”[1][3]

ComfyUI grew out of experiments Comfy began after discovering Stable Diffusion in October 2022. He recalled a background in web development, Python and basic automation, with no previous PyTorch programming or familiarity with node graph interfaces. He chose the graph because it seemed the best way to represent the diffusion process.[1]

Using AUTOMATIC1111, he explored high-resolution refinement: generate a low-resolution image, upscale it, then refine it. He modified the second pass's samplers and settings, then wanted to try a different model for that pass. He recalled deciding that building his own interface would be easier than extending the existing one. Working alone initially, he said he began coding on January 1, 2023, and released the first version on GitHub on January 16.[1]

In a Reddit demonstration, he paired Anythingv3 with Abyss Orange Mix 2 in his own GUI, explaining: “The first pass is really for composition and the second one is for detail.” He described combining Anythingv3's freedom with AOM2's style while overcoming what he considered AOM2's blurry backgrounds and overfitting. For character images, he advised limiting background tags in the first pass so they would not overwhelm the character description, then adding fuller background descriptions in the second.[3]

Another early experiment was area conditioning: apply different prompts to different image regions, then average the predictions at each diffusion step. Comfy recalled that posting this work on Reddit brought initial attention. He credited a video by Olivio, which he placed in March 2023, with a further burst of interest. As more people used ComfyUI, maintenance and feature requests began taking time away from his own experiments.[1]

Comfy said Stability hired him in June 2023 because it wanted to experiment with chaining SDXL's base and refiner models, and he worked on making SDXL run well in ComfyUI. In a historical Reddit discussion, he described ComfyUI as an internal testing tool and defended its flexibility for model experimentation. His criticism of AUTOMATIC1111 centered on its handling of the base-and-refiner workflow and its ability to accommodate future model changes.[1][4]

Comfy concentrated on the backend because he used ComfyUI locally himself. He attached an existing JavaScript node interface rather than spending substantial time building a frontend. His architectural preference was explicit: “Frontend and backend should be well separated with a defined API.” He considered Gradio useful for quickly demonstrating an ML project, but said he would avoid it for software intended to remain maintainable over a long period.[1]

GPU memory management was a central difficulty. In the 2025 interview, he described moving beyond completely loading and unloading models toward estimating the memory needed for sampling and removing as little already-loaded model data as possible. Retaining models avoided repeated loading, but exceeding GPU capacity could make Windows drivers page data into system RAM and slow execution sharply. He also described Python and PyTorch as limiting his fine-grained control over memory freeing.[1]

He sought to expose model controls while offering sampler nodes ranging from basic to advanced. His advice was to learn parameters through their visible effects on images, rather than requiring a complete theoretical understanding first. He also said he implemented Mochi in ComfyUI, explaining how its temporal VAE compressed time as well as space, unlike the spatial-only compression he described in earlier video-generation approaches.[1]

His model assessments emphasized artistic results as well as consistency. In the interview published January 2025, he suggested trying SD3.5 for more creative results and Flux for greater consistency. He argued that users adopted new models when improvements were substantial, and described much ComfyUI user evaluation as an aesthetic judgment of generated images.[1]

Comfy credited easy custom-node creation with helping the ecosystem grow, while acknowledging that node packs often duplicated functionality. He discussed bringing some of that functionality into core as a possible remedy. He separately credited the ComfyUI Manager maintainer with searching GitHub for new nodes and adding them to the manager; the registry effort was intended to improve that discovery workflow.[1]

In the 2025 interview, Comfy described himself as handling the core backend while others focused on the frontend, which he acknowledged had long been neglected. He explicitly distinguished closed-model API integrations from his own responsibilities. He welcomed other applications and startups using ComfyUI, reasoning that their use could attract people to the ecosystem even without direct contributions. A Krita plugin was an example he highlighted of others building on ComfyUI as a backend.[1]

A later repository-transition announcement, posted by crystal_alpine, described comfyanonymous as having long been the sole reviewer responsible for code quality. It announced a move from his personal account to Comfy-Org and plans to broaden review, community input and eventually open governance.[5]