ComfyUI

Last reviewed Aug 24, 2026 · 738 words

ComfyUI is the professional's interface to local image generation — a node graph where models, prompts, samplers, and post-processing wire together visually — and its famously steep first hour buys the thing no simpler UI offers: every image you make carries its complete recipe, and every technique on the internet arrives as a file you can load and run. Drag any ComfyUI-generated PNG onto the canvas and the entire workflow that made it reconstructs itself. That one property changes local image generation from knob-fiddling into something you can version, share, and reproduce.

VRAM decides your model family

Like local LLMs, image models are VRAM-bound first and everything else second. The working table:

ComfyUI's efficient memory management is a real advantage at the low end — its lazy loading and offloading run models that crash other UIs on the same card. The catalogue's 8 GB minimum system RAM is honest, but 32 GB is comfortable once models, VAE, and upscalers stack up. Models live in models/checkpoints, LoRAs in models/loras, VAEs in models/vae; they're multi-gigabyte each, so put the install on a roomy disk and expect a 100+ GB folder within months — the same disk-hygiene story as Ollama, with the same fix of periodically deleting the experiments.

Learn it in the right order

The default workflow (load checkpoint → prompt → KSampler → VAE decode → save) is seven nodes and teaches the entire mental model; run it, then change one thing at a time — sampler, steps, CFG — and watch what moves. Resist building graphs from scratch early: load other people's workflows (Comfy's template library, plus the PNG-drag trick on any image whose look you admire) and read them like annotated code. The graph that intimidates on day one reads fluently by day ten, and the payoff is that img2img, inpainting, ControlNet, and upscaling chains are just more nodes, not new applications to learn. If you only ever want prompt-in-image-out with no graph, AUTOMATIC1111 remains the friendlier tab-and-slider take — but it plateaus exactly where ComfyUI begins.

Custom nodes: the ecosystem and the risk, in one feature

ComfyUI-Manager is the app store: it installs the community nodes that workflows depend on and resolves the red "missing node" boxes automatically. Understand what you're doing when you click install: custom nodes are arbitrary Python executing with the app's permissions, published by strangers, and the ecosystem has had real malicious-package incidents. Sensible posture: install the widely-used essentials, be suspicious of a brand-new node a random workflow demands, prefer running the whole thing in a container, and — non-negotiable — never expose ComfyUI to the internet. --listen for LAN access is fine; the API and UI have no authentication, and "code execution as a service" describes both the app and several of its past CVEs. Tailscale covers the remote-use case, per the same logic as every unauthenticated homelab service.

Make outputs findable and workflows deliberate

Two habits separate a tool from a mess. First, filename prefixes: set filename_prefix per project (project/character-v2) so output/ stays navigable after a thousand generations. Second, save named workflow files (.json) at milestones rather than relying on the ever-mutating canvas — "portrait-pipeline-v3.json" in a git repo is reproducibility for free, and six months later it's the only way you'll remember how you achieved that look. The embedded-recipe PNGs cover accidental archaeology; saved workflows cover the deliberate kind.

What I'd do

ComfyUI on the GPU box (container preferred), models on the big disk, Manager plus a handful of proven custom nodes, LAN/tailnet access only, the default workflow mastered before any downloaded graph, one-change-at-a-time experiments, and named workflow files per project. Expect the first evening to feel like circuit design and the second week to feel like a superpower — that curve is the tool working as intended, and it's why the graph model has quietly become the local AI stack's standard for images.