Running Flux.1 on consumer hardware: a complete optimization guide

Flux.1, developed by Black Forest Labs, has taken the generative art community by storm. With its unmatched prompt adherence, hyper-realistic text rendering and detailed anatomy, it has surpassed commercial closed platforms. However, its massive scale (12 Billion parameters) makes it a beast to run on consumer graphics cards.

In this guide, we will walk through the exact steps to configure your local machine to render Flux.1 Schnell or Dev models using compressed weights without sacrificing visual fidelity.

Understanding the memory bottleneck

A standard unquantized Flux.1 Dev model requires over 24GB of memory to load the model weights, text encoders (CLIP and T5XXL) and VAE. If your system runs out of VRAM, your OS will offload layers to system RAM, resulting in generation times of 5 to 10 minutes per image instead of seconds. Our objective is to compress these weights to fit comfortably inside consumer cards (8GB - 16GB VRAM).

Step 1: Get the right quantized weights

Thanks to the open-source community, we have access to NF4 (NormalFloat 4) and GGUF quantizations. NF4 provides an excellent balance of speed and prompt adherence, while GGUF allows you to select specific bit-widths (Q4, Q5, Q8) depending on your hardware.

Step 2: Install ComfyUI and model files

ComfyUI is the node-based pipeline of choice for optimized local execution. Its native execution handles low-VRAM memory swapping much more efficiently than automatic web interfaces.

# 1. Clone the ComfyUI repository
git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI

# 2. Install dependencies (make sure you have PyTorch with CUDA support)
pip install -r requirements.txt

Place your downloaded flux1-schnell-nv4.safetensors into the ComfyUI/models/unet/ folder and your text encoders (T5XXL and CLIP) inside ComfyUI/models/clip/.

Step 3: Leverage smart memory offloading

When running ComfyUI, launch it using command-line arguments that instruct the engine to stream weights from host RAM to VRAM dynamically:

python main.py --lowvram --use-split-cross-attention

The --lowvram flag forces ComfyUI to load the text encoders, perform the prompt-token parsing in system RAM, unload them and then load the image generation UNet model into VRAM. This process ensures that the text encoders and the generator model do not fight for the same GPU memory space simultaneously.