In this episode of the fxpodcast, we speak to ComfyUI co-founder, Yannik Marek (ComfyAnonymous) about how he got started, where this key software is heading.
There’s a quiet but very real shift underway in how AI is being integrated into production VFX pipelines, and tools like ComfyUI are increasingly at the center of that transition. What began as an experimental, node-based interface for generative workflows is rapidly evolving into something far more consequential: a flexible orchestration layer where artists, TDs, and researchers can begin to stitch AI into real, production.
Yannik Marek is the original creator of ComfyUI, which he began in early 2023 as a solo experiment with Stable Diffusion, and he built the first version in just a couple of weeks. With a background not in machine learning but as a C++ programmer working in his father’s dental implant factory.
Marek approached AI tools from a practical, systems-oriented perspective. His key innovation was designing a node-based, flowchart-style interface that allowed users to control different diffusion models at different stages of image generation, bringing a level of modularity and transparency that resonated strongly with technical artists. What started as a zero-funding open-source project has since grown rapidly, with ComfyUI reaching over 4 million users, around 150,000 daily downloads by early 2026, and a thriving ecosystem of more than 60,000 community-built nodes.
There’s a growing sense that AI in VFX is finally moving beyond isolated demos and into something that resembles a usable VFX pipeline, and much of that momentum is being driven by ComfyUI. The platform is quickly emerging as a kind of connective tissue for AI-driven workflows, offering artists a node-based environment that feels closer to established VFX tools than most generative interfaces. Similar steps forward have been seen with the Foundry buying GripTape (see our Story here)
Within the Compfy ecosystem, recent developments such as the Radiance plugin from FXTD Studios highlight just how far this ecosystem is evolving: unlike standard nodes that clamp to 8-bit, Radiance preserves full IEEE 754 floating-point precision across the entire chain, eliminating banding and data loss, and introducing a level of fidelity that aligns with professional VFX standards.
We will be covering more on Radiance in the coming weeks here on fxguide, but with features like the Radiance Workspace node handling project paths, versioning, and state management for 32-bit data, it points to a broader shift, namely that AI tools that are not just generative, but editable, trackable, and pipeline-aware. Radiance itself is only one example, but it underscores a larger trend: ComfyUI is becoming an increasingly important foundation for integrating AI into real, production-ready VFX workflows.
In the fxpodcast
In the interview, Yannik Marek explains that ComfyUI began as a highly practical experiment, born out of frustration with early Stable Diffusion interfaces that limited resolution and flexibility. Wanting finer control over multi-stage image generation (such as using different models or samplers across passes), Marek built a node-based system that allowed modular experimentation, an approach that has since become central to ComfyUI’s identity. He emphasises that development today is largely driven by rapid advances in new models, with ComfyUI providing “day zero” support through partnerships with leading ML labs, while the broader community builds on top of this foundation through an enormous ecosystem of custom nodes.
The discussion highlights how this openness creates both strength and tension. Community contributions and partner nodes (including closed-source APIs) expand capabilities quickly but introduce challenges around vetting, stability, and security—particularly in production environments. Marek notes that while there is some light vetting, especially for more complex open-source integrations, much of the ecosystem remains intentionally flexible rather than tightly controlled.
To address consistency and enterprise concerns, newer initiatives such as Comfy Cloud provide a more controlled, standardised environment (same hardware, same stack), though full version “freezing” is not yet implemented. Meanwhile, App Mode reflects another key shift, making workflows accessible beyond technical users by wrapping node graphs into simple, app-like interfaces. Taken together, the interview paints ComfyUI as a fast-evolving platform balancing experimentation and production needs: driven by community energy, shaped by rapid model innovation, and increasingly extending toward more stable, shareable, and pipeline-aware AI workflows.