ComfyUI Guide for Beginners (2026): Install, Models, Nodes, Settings and Fixes

ComfyUI is the most powerful free way to run image, video and audio AI models on your own PC, and also the most confusing one to start with. You get a canvas of boxes and wires, a models folder with fifteen subfolders, and an error log that only makes sense once you know where to look. I have written a separate deep guide for each part of it. This page puts them in the order you need them, gives you the short answer for each step, and links to the full walkthrough when you want the details.

What Is ComfyUI?

ComfyUI is a free, open-source app that runs diffusion and other generative models through a node graph. Each box (node) does one job: load a model, encode a prompt, sample, decode, save. You wire them together into a workflow and press Run. It runs in your browser, but everything happens locally on your GPU.

Why people put up with the learning curve:

• New models land here first. FLUX, Qwen-Image, Wan, LTX, MOSS-Image and MiniMax H3 usually get native ComfyUI support on day one or within a week.
• You see and control every step. Swap a loader, add a LoRA, upscale, inpaint or chain image-to-video in one graph.
• Workflows are portable. A PNG made in ComfyUI carries its whole workflow, so you can drag someone’s image in and get their exact setup.
• It is free and private. No credits, no uploads, and it works offline once the models are downloaded.

What it is not: a one-button app. If you want a simple chat-style image generator, the cloud tools are easier. If you want control and the newest open models, ComfyUI is worth an afternoon.

The ComfyUI Roadmap: Which Guide Do You Need?

Step 1: Install ComfyUI

Use Python 3.13 and, on NVIDIA, PyTorch built for CUDA 13.0 (cu130) with a current driver. Full steps for every option are in my ComfyUI install guide. RTX 50 owners should also read the RTX 50 setup guide, mainly to avoid installing xformers, which can pull in the wrong PyTorch.

Step 2: Check Your GPU and RAM

System RAM matters more than people expect, because ComfyUI parks models in RAM when it swaps them out of VRAM. Treat 32 GB as the minimum and 64 GB as the comfortable size for video. If the terminal closes or prints Killed while a model loads, you ran out of RAM, not VRAM. See how much RAM you need and, for two cards, the multi-GPU guide.

Step 3: Put Models in the Right Folders

Most “model not found” problems are a file in the wrong folder. Modern models are split into three parts, each with its own folder and loader.

After copying a file, press R in ComfyUI to refresh the dropdowns. Subfolders are fine; they show up as a path in the menu. If you run several installs, share one models folder with extra_model_paths.yaml instead of copying 200 GB around. The full list is in the model folder cheat sheet.

Step 4: Make Your First Image

1. Open http://127.0.0.1:8188 in your browser.
2. Go to Workflow > Browse Templates and pick a text-to-image template for a model you have (or that the template offers to download).
3. Check each loader points at a file you actually have. A red outline or “Value not in list” means a missing file.
4. Type a prompt in the positive text box. Keep the template’s steps, CFG and sampler for your first run.
5. Click Run and watch the terminal: it shows each model loading and the sampling progress.
6. Find the result in ComfyUI/output/ or the queue sidebar.

The key nodes you will see in almost every image workflow are: a model loader, a text encoder (CLIP Text Encode), an empty latent (your image size), KSampler, VAE Decode and Save Image. Once you know those six, most workflows stop looking scary.

Step 5: Add the Manager and Custom Nodes

Custom nodes add features ComfyUI does not ship with: video loading, GGUF support, face detailers, live VRAM monitors and more. The Manager installs and updates them.

• Desktop: the Manager is already on.
• Portable: .\python_embeded\python.exe -m pip install -r ComfyUI\manager_requirements.txt, then add --enable-manager to run_nvidia_gpu.bat.
• Manual: pip install -r manager_requirements.txt, then python main.py --enable-manager.

When you load a workflow with red nodes, click Install All in the missing-nodes prompt, then restart ComfyUI fully. My first five installs on any new setup are rgthree-comfy, ComfyUI-VideoHelperSuite, ComfyUI-KJNodes, ComfyUI-Crystools and ComfyUI-GGUF.

Safety: custom nodes are Python code that runs on your PC. Keep the Manager security level on normal, prefer popular maintained packs, and never set it to weak on a server reachable from the internet. Details: Manager guide and must-have custom nodes.

Step 6: Samplers, Flags and Speed

CFG above 1 doubles the work per step, which is why distilled 4-8 step models use CFG 1. “Ancestral” and “SDE” samplers add noise every step, so the image keeps changing as you change the step count. More in samplers and schedulers explained.

Launch flags go at the end of the python main.py line or in run_nvidia_gpu.bat. The useful ones:

• --use-sage-attention: faster attention once SageAttention is installed (about 30-35% on video). Install guide; never use plain pip install sageattention on Windows.
• --reserve-vram 1.5: leaves room for your browser and other apps, which stops many random OOMs.
• --listen: open ComfyUI from another device on your network.
• --preview-method auto: live previews while sampling.
• --disable-all-custom-nodes: the fastest way to test whether a node broke ComfyUI.

Modern ComfyUI manages VRAM on its own, so --lowvram does little on most NVIDIA cards now. The full list is in the command line arguments cheat sheet.

Step 7: Run Big Models on Small GPUs

Smaller files beat any flag. In order of what helps most:

1. Use fp8 or INT8 files where the model offers them. They load with the normal nodes and run fast on RTX 40/50.
2. Use GGUF when even fp8 does not fit: install ComfyUI-GGUF, swap Load Diffusion Model for Unet Loader (GGUF). Q8_0 is close to full quality, Q5_K_S or Q4_K_S for 8-12 GB. GGUF guide
3. Shrink the text encoder too. A GGUF T5 or Qwen encoder saves several GB with almost no visible change.
4. Lower resolution or frame count, and use VAE Decode (Tiled) for big images and video.
5. On RTX 50, try NVFP4 models: one published FLUX.1 Dev test showed 7.73 it/s vs 4.21 it/s for FP8 on an RTX 5090.

For MiniMax H3, Krea 2, Ideogram 4 and Qwen-Image 2.1 GGUFs, use the leejet/ComfyUI-GGUF fork rather than city96’s original, and never install both. If you still hit OOM, the CUDA out of memory fix walks through the rest.

Step 8: Save, Share and Automate Workflows

• Save (Ctrl+S) keeps the workflow inside ComfyUI, in ComfyUI/user/default/workflows/.
• Export downloads a .json you can back up or share.
• The PNG is the workflow. Images from Save Image carry the full graph in their metadata; drag one onto the canvas to load it. Discord, X and most chat apps strip that metadata, so share the JSON instead.
• Export (API) gives the compact format for scripts. Send it to POST /prompt and listen on the WebSocket for progress and images.

Guides: save, export and share workflows and the ComfyUI API from Python.

Step 9: Update Without Breaking Things

Take a Manager snapshot first, update custom nodes after ComfyUI, and restart fully. If an update breaks something, restore the snapshot. New models usually need the latest build, so if a template will not load or a model says “unknown architecture”, updating is the first fix. Full steps, rollback and uninstall: how to update ComfyUI safely.

What to Build Next

Once the basics work, these hubs and guides take you further:

• Image to video: ComfyUI image to video guide (Wan, LTX and more).
• MiniMax H3 video with audio: MiniMax H3 ComfyUI guide.
• MOSS-Image, text and layout control: MOSS-Image in ComfyUI.
• Local LLMs next to ComfyUI: the complete Ollama guide.

Common Problems and Fixes

The full list of 22 errors with causes is in ComfyUI errors and fixes, and the browser-side issues are in the reconnecting fix.

Frequently Asked Questions

Is ComfyUI free?

Yes. ComfyUI is free and open source, and running models locally costs nothing beyond electricity. Comfy also sells optional cloud and API features, which you do not need.

Is ComfyUI hard to learn?

The first hour is the hardest. Start from the built-in templates instead of an empty canvas, learn the six core nodes, and it gets much easier quickly.

What GPU do I need for ComfyUI?

An NVIDIA card with 8 GB of VRAM runs SDXL and small GGUF quants of modern models. 12-16 GB is the sweet spot for FLUX, Qwen-Image and video with fp8 or GGUF. AMD works through ROCm, and Macs work but are slower.

Can I run ComfyUI without a GPU?

It runs on CPU with --cpu, but it is very slow. A rented cloud GPU such as RunPod is a better option if you have no graphics card.

Where does ComfyUI save images?

In ComfyUI/output/ by default. Saved workflows go in ComfyUI/user/default/workflows/.

Is ComfyUI safe?

ComfyUI itself is safe. The risk is custom nodes, which run code on your PC, so install popular maintained packs and keep the Manager security level on normal.

ComfyUI Desktop or portable?

Desktop is easier and updates itself. Portable is a single folder you fully control and can copy as a backup. Both run the same ComfyUI.

Final Thoughts

Install it, put three model files in the right folders, run a template and get one good image. Then add the Manager, a GGUF loader and SageAttention, and you can run almost any new open model the week it comes out. When something breaks, the terminal almost always tells you why. I keep this page updated as I add ComfyUI guides, so bookmark it.

Sources

• Comfy-Org/ComfyUI on GitHub and ComfyUI documentation
• ComfyUI docs: Manager installation
• ComfyUI docs: How to update ComfyUI
• city96/ComfyUI-GGUF