ComfyUI

ComfyUI is a local diffusion engine with a graph/node interface: instead of typing a prompt into a box, you wire nodes together — model loader, text encode, sampler, decoder, save — and the graph is the pipeline. It supports image and video generation models locally and has become the de-facto standard for reproducible, shareable generation workflows.

Why it is on our list

• It is the local image/video path that actually works offline. No credits, no queue, no content filter you did not configure.
• Workflows are files. A graph can be exported, shared, version-controlled, and reviewed — which is the same property we value in Markdown knowledge.
• Enormous ecosystem. Custom nodes extend it to video, upscaling, inpainting, control nets, and more.

Install and first run

1. Download from https://comfy.org/download. Comfy Desktop is the recommended path for most people; the Windows portable build and manual git installs exist for people who need the latest commits or finer control.
2. Launch, then open an example workflow from the templates rather than starting from an empty canvas.
3. Run a simple text-to-image graph end to end before installing any custom nodes.

Official references:

• System requirements
• First generation walkthrough

Configuration worth knowing

• Model placement. Checkpoints, VAEs, LoRAs, and upscalers go in their own subfolders; a model in the wrong folder is the classic "ComfyUI cannot see my model" cause.
• Desktop vs portable vs git. Desktop tracks stable releases and supports snapshots/rollback; portable tracks the stable release standalone; git installs can track the latest commits.
• Update channel. Comfy Desktop lets you choose between stable and latest GitHub — relevant when a workflow depends on a feature that has not reached stable.

Common pitfalls

• Installing custom nodes before the base install works. If a plain text-to-image graph fails, no custom node will save you.
• "Install missing nodes" does not always work. Workflow authors share graphs that depend on node packs; missing-node resolution is a frequent source of breakage, and it is the single most common complaint from beginners.
• VRAM exhaustion on video models. Image models that fit comfortably are far smaller than video pipelines. Start with images.
• Disk bloat. Model files are multi-gigabyte; a working install can quietly reach hundreds of gigabytes.
• Licence drift in custom nodes. The engine is GPL-3.0, but each custom node pack carries its own licence. Check before shipping a workflow inside a product.
• Confusing the desktop app with the engine. They are separate repositories and ship on different schedules.

Tips and tricks

• Learn the graph by editing an example workflow, not by building from zero.
• Snapshot or back up a working install before bulk-updating; one bad node pack can break a graph you depend on.
• Treat exported workflows like code: keep them in git alongside a note of the model filenames and node pack versions they need.
• If you only want occasional images, a smaller checkpoint with a proven workflow beats a huge model with an unloved graph.

Videos

Verified via YouTube oEmbed on 2026-10-28.

• ComfyUI Course — Learn ComfyUI From Scratch | Full 5 Hour Course (Ep01) — pixaroma.
• ComfyUI in 20 Minutes: Absolute Beginner Guide to Getting Started — WINBUSH.
• ComfyUI Tutorial Series: Ep01 — Introduction and Installation — pixaroma.
• Ultimate Beginner Guide to Learning ComfyUI (2026) — Max Novak.

Related

• Unsloth Desktop — also handles diffusion models, if you want one app for everything.
• Choosing local AI tools — is image generation your goal, or a side quest?
• Docker Desktop — for the containerised install path.

Verification note: licence and repository from the GitHub API; install options, update channels, and requirements from https://docs.comfy.org on 2026-10-28. The repository moved to the Comfy-Org organisation — older links to comfyanonymous/ComfyUI redirect.