What for: a node-based editor for generating images and video locally — maximum control over the whole pipeline, and free.

What it is and who it's for

ComfyUI is a node-graph editor for generating images and video locally: you assemble a pipeline out of nodes (model to prompt to sampler to post-processing) and get maximum control. It's for advanced users and artists who need reproducibility, fine-grained tuning and support for any open model (SDXL, FLUX, video) — free and private. It stands out for unmatched pipeline flexibility and a huge ecosystem of custom nodes.

Key features

Visual graph pipeline: full control over every step of generation.
Any open model: SDXL, FLUX, video models, ControlNet, LoRA, inpainting.
Reproducible workflows (json) — share and reuse them.
ComfyUI Manager and thousands of community custom nodes.
Local and private; no limits and no per-generation billing.
Batch processing and complex multi-stage pipelines.

Get started in 5 minutes

Install it (the desktop app or git are the recommended routes):

git clone https://github.com/comfyanonymous/ComfyUI && cd ComfyUI && pip install -r requirements.txt && python main.py

Open http://localhost:8188 and drop a model checkpoint into models/checkpoints.
Grab a ready-made workflow (drag and drop the json) or build your own from nodes, then hit Queue.

When to use it, when not to

✅ Use it if you want maximum control, reproducibility and any open model running locally for free.
✅ Use it if you're building complex pipelines (ControlNet, LoRA, video, upscale chains).
❌ Skip it — go with "Leonardo AI" — if you want similar control but in the cloud, without local fiddling.
❌ Skip it — go with "Midjourney"/"Nano Banana Pro" — if you want fast results without wiring up a graph.

The honest price

Completely free and open source. Your costs are a powerful GPU (VRAM) and the time to learn it. You download models and nodes yourself. Exact requirements depend on your pipeline.

Gotchas

High barrier to entry — graphs scare beginners off.
You need a strong GPU; heavy pipelines are VRAM-hungry.
Custom nodes off the internet are a supply-chain risk; install from trusted sources only.
Node and model compatibility sometimes breaks on updates.

🤖 Prompt accelerator

"I'm learning ComfyUI on with a <MODEL/VRAM> GPU. Help me build a workflow for <DESCRIBE: txt2img on FLUX / inpaint / pose ControlNet / upscale>: which nodes I need, how to wire them, which models and LoRAs to download and where to put them. Give me a step-by-step graph build and the best sampler settings for my VRAM.