From Nodes to Album Covers: Joshua Davis’s ComfyUI Course

Every AI art conversation eventually lands on the same anxious question: is this tool creating, or just copying? "Working with ComfyUI," a three-week course taught by digital art pioneer Joshua Davis, was built to answer that question with a workflow instead of a debate. Students didn't type prompts and hope. They built node-based systems from their own images, models, and ideas, then used that system to completely reimagine a music artist's next album cover.

This article breaks down what ComfyUI actually is, who taught the course, how the class was structured, and what students walked away with, including a few final projects that took the brief further than anyone expected.

ComfyUI is an open-source, node-based interface for building AI image and video generation pipelines. Instead of typing a prompt into a box and hoping for the best, users construct a visual graph: each node represents a discrete step, loading a checkpoint, encoding a prompt, sampling noise, decoding the final image, and every node is wired together by hand. The result is full visibility and full control over the entire generation process, not just the output.

That transparency is exactly why ComfyUI has become the standard for serious AI creators in 2026. It runs every major diffusion model released since 2023, Stable Diffusion, SDXL, Flux, and beyond, without forcing users to learn a new tool every time a new model drops. It runs locally, keeping data private and workflows fully reproducible. And its open-source ecosystem means thousands of community-built nodes are available for anything from background removal to relighting to style transfer.

In short: ComfyUI turns AI image generation from a black box into a workbench. That's precisely the mindset the course was designed around.

You don't teach a course like this without a serious track record, and Joshua Davis has one. Since 1995, Davis has built a career as a designer, technologist, and new media artist by writing his own code, generating rule-based, randomized visual compositions rather than relying on pre-built tools.

His work speaks for itself. Davis has collaborated with the UEFA Champions League, Adidas, and Pepsi x Pharrell Williams. His pieces sit in the permanent collections of the Whitney Museum of American Art and the Smithsonian's Cooper Hewitt Design Museum, and he's shared his process on stages like TED and 99U.

That background shaped how the course was taught. Rather than treating ComfyUI as a black-box shortcut, Davis walked students through his own creative process session by session, showing real use cases, not just theory, and demystifying the node-based system from the ground up.

"Working with ComfyUI" ran for three weeks (45 hours total), built around a simple but firm principle: AI doesn't have to mean stealing other people's work. Instead of relying on scraped datasets or someone else's imagery, students learned to build workflows powered by their own assets, turning ComfyUI into a rapid-prototyping tool rather than a replacement for creative judgment.

The curriculum moved from fundamentals to full production:

• Foundations — Stable Diffusion and ComfyUI setup, inputs, outputs, models, and LoRAs

• Core mechanics — KSampler, seeds, steps, CFG, samplers, and schedulers

• Composition and style control — ControlNet for composition transfer, IPAdapter for style transfer

• Scene building — background removal and positioning with BiRefNet and LayerUtility, relighting with IC-Light

• Advanced generation — Flux, Florence2, and Searge LLM

• Post-production — refining outputs with traditional image editing techniques

Notably, the course never treated ComfyUI as the only tool in the room. Students were encouraged to mix in whatever got the job done, several brought ChatGPT and Claude into their concepting process, and finished projects in Figma or Photoshop. The point was never tool purity. It was building a workflow that actually delivers the vision in your head.

For the final project, the brief was intentionally open: pick a music artist and redesign their next album, and go as far as your creativity would take you. Some students stayed close to the brief. Others took "as far as your creativity takes you" as a personal challenge, pushing full concept album visuals, alternate covers, and campaign-style variations built entirely from their own ComfyUI pipelines.

Every final presentation opened the same way, too: with donuts. It became a quiet tradition of the class, a reminder that behind every carefully engineered node graph was still a room of people having fun making things.

The bigger lesson here isn't really about nodes, samplers, or checkpoints. It's about ownership. In a moment where "AI art" often means typing a prompt and hoping it doesn't resemble someone else's copyrighted work too closely, this course modeled a different approach: use AI as a rapid-prototyping engine for your own ideas and your own assets, then finish the work with the same post-production rigor any professional creative would use.

That's a workflow, not a shortcut, and it's exactly the kind of AI literacy more design and art programs are going to need.