How to Run Flux 3 Image Klein Locally in ComfyUI: 8GB VRAM Setup Guide

Why Local Flux 3 Image Klein is a Game Changer

When Black Forest Labs dropped FLUX 3 Image Klein on October 15, 2026, the biggest surprise was its hardware accessibility. While the flagship FLUX 3 model demands data-center-grade compute (often exceeding 24GB–32GB VRAM), Klein is a streamlined 5.6-billion-parameter model.

Thanks to INT8 and FP8 quantization builds, you can now run state-of-the-art flow-matching text-to-image and inpainting on consumer NVIDIA GeForce RTX 3060, 4060, and 4070 graphics cards with just 8 GB of VRAM.

In this guide, we break down the exact ComfyUI setup, required nodes, model weight links, and inference speed expectations.

Minimum & Recommended Hardware Specifications

Step 1: Update ComfyUI & Core Dependencies

Ensure your ComfyUI installation is updated to the latest commit supporting the new Qwen3-VL text encoder and flow-matching schedulers:

cd ComfyUI
git pull
pip install -r requirements.txt --upgrade

Step 2: Download the Official Model Weights

Download the official open weights from the Hugging Face repository (black-forest-labs/flux-3-image-klein):

1. Diffusion Model: Download flux-3-image-klein-fp8.safetensors (approx. 5.6 GB) or the INT8 quantized build into ComfyUI/models/unet/.
2. Text Encoder: Download the dedicated qwen3_vl_4b_instruct.safetensors into ComfyUI/models/clip/.
3. VAE: Download the standard FLUX VAE into ComfyUI/models/vae/.

Step 3: Configure the 8GB Low-VRAM Generation Workflow

1. Open ComfyUI in your browser (http://127.0.0.1:8188).
2. Add the UNETLoader and point to your Klein safetensors file.
3. In the execution flags, pass --lowvram or --fp8_e4m3fn-text-enc if running on an 8GB card to prevent CUDA out-of-memory errors during the initial text tokenization pass.
4. Set the sampling steps to 16 steps (Klein is step-distilled and requires no CFG guidance scale, meaning CFG should be locked at 1.0).

Step 4: Speed and Performance Benchmarks

• RTX 5090: ~4.1 seconds per 1024x1024 image.
• RTX 4070: ~8.8 seconds per 1024x1024 image.
• RTX 3060 (8GB): ~14.5 seconds per 1024x1024 image.

Don't Want to Deal with Local Setup Errors?

Running locally is great for offline experimentation, but managing Python dependencies, CUDA toolkit mismatches, and thermal throttling can be tedious.

With Flux 3 Image Klein Online, you can generate native 4K visuals instantly in your browser powered by cloud clusters—zero installation, zero VRAM limits, and full commercial usage rights included.