ComfyUI Fix, Crash & Optimization Guide

ComfyUI is a node-based graphical interface for Stable Diffusion and other diffusion models, run locally through Python. It is used by artists and ML practitioners who want fine-grained control over image-generation workflows without writing code.

Install / First Setup

ComfyUI is distributed as a Python project. The recommended path is a git clone plus a dedicated virtual environment:

git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
python -m venv venv
# Windows
venv\Scripts\activate
# macOS / Linux
source venv/bin/activate
pip install -r requirements.txt

Launch (default web UI on http://127.0.0.1:8188):

python main.py

On Windows, the portable package bundles Python and is the simplest starting point. For GPU acceleration you need a working NVIDIA driver and a CUDA-enabled PyTorch build.

Common Issues & Fixes

Torch not compiled with CUDA enabled / no GPU available

Cause: A CPU-only PyTorch wheel was installed, or no compatible GPU/driver is present. Fix: Reinstall PyTorch with the CUDA build matching your driver. Check with python -c "import torch; print(torch.cuda.is_available())". If it returns False, reinstall the CUDA wheel from the official PyTorch site rather than the default pip torch.

Out of VRAM during generation

Cause: The model, VAE, or attention buffers exceed available GPU memory. Fix: Launch with memory-saving flags:

python main.py --lowvram
# or, less aggressive
python main.py --medvram

You can also reserve headroom with --reserve-vram 1.5 (GB) and reduce resolution or batch size in the workflow. For VAE decode OOM, enable tiled VAE decoding in the VAE Decode node settings.

Black or blank generated images

Cause: Missing or mismatched VAE, or a silent VAE decode out-of-memory. Fix: Ensure a compatible VAE is loaded in the VAE Loader node. If images are black only on large outputs, switch the VAE Decode to tiled mode to avoid VRAM spikes.

Cannot import custom node / missing module

Cause: A custom node's dependencies were not installed. Fix: Each custom node typically ships a requirements.txt. Install it inside the same venv, then restart python main.py. Use the ComfyUI-Manager to install and update nodes from the UI.

Web UI won't open / port already in use

Cause: Another ComfyUI instance or process holds port 8188. Fix: Either stop the conflicting process or change the port: python main.py --port 8189. To expose on your LAN use python main.py --listen 0.0.0.0 (only on trusted networks).

Performance & Optimization

• Low-End (8 GB system RAM, no/weak GPU): Run with --cpu or --lowvram, use SD1.5-class models (1–2 GB), avoid SDXL. Expect slow, single-image generation.
• Mid-Range (16 GB RAM, 6–8 GB VRAM GPU such as RTX 3060): Use --medvram. SDXL (base + refiner) fits at modest resolution (e.g. 1024×1024). Keep batch size at 1.
• Workstation (24+ GB VRAM GPU): Run without VRAM flags. Enable xformers or the default SDPA attention for speed, raise batch size, and use fp16 weights to roughly halve model memory versus fp32.
• Use --use-pytorch-cross-attention or xformers only when compatible with your CUDA version; mismatches cause import errors.

Version & Compatibility Notes

ComfyUI updates frequently via git pull and does not use fixed release numbers. It targets Python 3.10+ and PyTorch 2.x. xformers is optional and version-sensitive; if it fails to import, run without it. For exact supported CUDA/PyTorch combinations, consult the official ComfyUI release notes and the PyTorch download page.

FAQ

Q: How do I update ComfyUI safely? A: Run git pull inside the ComfyUI folder, then update custom nodes through ComfyUI-Manager. Restart python main.py afterward.

Q: Can ComfyUI run entirely on CPU? A: Yes, with python main.py --cpu, but generation is dramatically slower. It is usable only for small SD1.5 models.

Q: What does --lowvram actually do? A: It offloads model weights to system RAM and loads them to VRAM in chunks, trading speed for the ability to run larger models on limited GPUs.

Q: Where do generated images go? A: They are saved in the ComfyUI/output directory by default, or wherever your Save Image node points.

Q: How do I load a GGUF or quantized model? A: Install the ComfyUI-GGUF custom node and use its UNET/CLIP/VAE loader nodes, selecting the GGUF file from your models folder.

Q: Why is the UI slow to load the first time? A: The first launch compiles/imports many modules and may download default model files. Subsequent starts are faster.

Related Guides

• vLLM Fix & Optimization Guide
• Ollama Fix & Optimization Guide
• Automatic1111 Fix & Optimization Guide
• Dev RAM Calculator

Accuracy Note

Commands and paths reflect common, real-world setups as of 2026-08. Always verify against your installed version and OS. When in doubt, consult the official ComfyUI documentation.