ComfyUI is the most powerful node-based interface for running local AI image and video generation. This guide gets you from zero to a working workflow on Windows with an NVIDIA GPU — tested live on an RTX 5080 and RTX 3080 16GB.
Hardware Requirements
| GPU | VRAM | Best Use Case |
|---|---|---|
| RTX 5080 | 16GB | Full quality, all models, fast batch |
| RTX 4090 | 24GB | Full quality, FLUX FP16, large batch |
| RTX 3080 16GB | 16GB | Full quality, slower on large models |
| RTX 3080 10GB | 10GB | Reduced resolution, GGUF models |
| RTX 3060 | 12GB | Standard models, SD1.5 and SDXL |
| GTX 1660 Ti | 6GB | SD1.5 only, very slow |
Step 1 — Install Prerequisites
You need three things before cloning ComfyUI: Python 3.10, Git, and the CUDA Toolkit.
Python 3.10.x
Download from python.org — use exactly 3.10.x, not 3.11 or 3.12. Many custom nodes have dependency conflicts with newer versions.
python --version
Git
git --version
CUDA Toolkit
Download CUDA 12.1 from [nvidia.com/cuda-downloads](https://developer.nvidia.com/cuda-downloads). Match your driver version.
nvcc --version
Step 2 — Clone ComfyUI
cd C:\ git clone https://github.com/comfyanonymous/ComfyUI cd ComfyUI
Step 3 — Install Dependencies
python -m venv venv.\venv\Scripts\activate
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt
The PyTorch install will download ~2.5GB. This is the step most people mess up — make sure you're using the cu121 index URL, not the default PyPI version.
Step 4 — Install ComfyUI Manager
ComfyUI Manager adds a one-click node installer directly inside the UI. It's essential.
cd custom_nodes git clone https://github.com/ltdrdata/ComfyUI-Manager cd ..
After this, every time you find a workflow that needs missing nodes, ComfyUI Manager will detect and install them automatically.
Step 5 — Download Your First Model
Place models in ComfyUI/models/checkpoints/
For FLUX (requires 12GB+ VRAM):
Step 6 — Launch ComfyUI
python main.pypython main.py --gpu-only --highvram
python main.py --gpu-only
python main.py --lowvram
Open your browser to: http://127.0.0.1:8188
127.0.0.1:8188 — it's faster than typing it each time.
RTX 5080 Optimal Settings
These settings max out quality on 16GB VRAM with headroom for LoRAs:
Model: FLUX Dev FP8 or SDXL Batch size: 2–4 images Resolution: 1024x1024 (SDXL) / 768x512 (video) Steps: 20–30 CFG Scale: 7.0 (images) / 3.0 (video) Sampler: dpmpp_2m Scheduler: karras VAE: Built-in (baked into checkpoint) Precision: fp16
Installing Essential Custom Nodes
In ComfyUI Manager, search and install these first:
- 1ComfyUI-Impact-Pack — face detailing, segmentation, upscaling
- 2ComfyUI_IPAdapter_plus — image prompt control
- 3ComfyUI-AnimateDiff-Evolved — video animation
- 4ComfyUI-VideoHelperSuite — video input/output (required for LTX)
- 5rgthree-comfy — better node organization and groups
Your First Workflow
- 1Click Load in the top menu
- 2Select
default_workflow.json - 3In the
Load Checkpointnode, select your downloaded model - 4Click Queue Prompt
Performance Benchmarks
Tested on RTX 5080 16GB, --gpu-only --highvram:
| Model | Resolution | Steps | Time |
|---|---|---|---|
| SDXL FP16 | 1024×1024 | 20 | ~3.2s |
| FLUX Dev FP8 | 1024×1024 | 20 | ~8.1s |
| SD 1.5 | 512×512 | 20 | ~0.8s |
| LTX Video 2.3 | 768×512 97f | 25 | ~5.4s |
Common Issues
"CUDA out of memory" — Add--lowvram to launch command, reduce batch size to 1, or switch to a GGUF quantized model.
Black images — Your VAE doesn't match the model. Download the correct VAE for your checkpoint and load it manually with a VAELoader node.
Slow generation — Check --gpu-only flag is set and your venv is activated. Without the venv, you'll use the system Python and possibly CPU inference.
Nodes missing — Open ComfyUI Manager → click "Install Missing Custom Nodes". This auto-detects and installs everything a workflow needs.
You're ready to run your first workflow. Next up — training your own LoRA to lock a character or style into any model.