Setup guide / Flux 2 Klein Base 9B

Get ComfyUI running.

This is the clean install path for ALTERA2ION exterior and interior packages. It is built around the package-ready local baseline plus the browser-first Comfy Cloud option.

4

Setup paths

22GB+

Local minimum

24GB

Practical GPU class

ComfyUI official install path FLUX 2 Klein Base 9B package baseline 22GB minimum / 24GB-class NVIDIA GPU No local GPU? Use Comfy Cloud

Windows / local install

Before you start.

Official ComfyUI currently recommends the Windows portable build as the fastest local path. For ALTERA2ION, the important part is launching cleanly and placing the model stack in the right folders before you open the package JSON.

Component Minimum for ALTERA2ION Notes
GPU 24GB-class NVIDIA GPU RTX 3090, RTX 4090, or equivalent.
VRAM 22GB minimum BFL currently lists Base 9B at 21.7GB VRAM, so 24GB cards are the practical local floor.
RAM 32GB system memory 64GB is the cleaner daily-use target.
Storage 100GB free SSD High-speed local storage keeps model loads cleaner.
OS Windows 10/11 or Linux Use the local ComfyUI path that matches your machine.
ComfyUI Releases ComfyUI GitHub Official Docs Python

Windows / first run

Install, place, run.

01 / Recommended path

Use the official Windows portable build.

Download the current portable release, extract it, then run it once before adding the ALTERA2ION files.

02 / Required package files

Place the model stack in the right folders.

These package workflows do not use a single checkpoints folder. They rely on a diffusion model, text encoder, VAE, and the package LoRA in their matching ComfyUI directories.

ComfyUI\models\diffusion_models\flux-2-klein-base-9b-fp8.safetensors
ComfyUI\models\text_encoders\qwen_3_8b_fp8mixed.safetensors
ComfyUI\models\vae\flux2-vae.safetensors
ComfyUI\models\loras\[your-package-lora].safetensors

03 / Model folders

Place the base model, text encoder, VAE, and package LoRA in the matching folders before you load the supplied workflow.

04 / Load workflow

Import the supplied JSON and keep the packaged settings unchanged for the first run.

05 / Launch

Load your base image and your reference image, then queue the graph.

Alternative install

If you prefer the official CLI route instead of the portable build, ComfyUI documents the following bootstrap path:

pip install comfy-cli
comfy install

Linux / local install

Manual, stable, clean.

Linux is the cleanest sustained local path if you already run a dedicated workstation. The official ComfyUI guidance currently calls Python 3.13 very well supported, with 3.12 as the fallback if a custom node is fussy.

01 / Python + repo

sudo apt update
sudo apt install -y python3.13 python3.13-venv python3-pip git
git clone https://github.com/Comfy-Org/ComfyUI ~/ComfyUI
cd ~/ComfyUI
python3.13 -m venv venv
source venv/bin/activate

02 / NVIDIA PyTorch

pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130
pip install -r requirements.txt

03 / Model stack + manager

After the base install, place the same four files used on Windows into the matching model folders, then run ComfyUI and load the supplied JSON.

python main.py --enable-manager
ComfyUI GitHub Official Docs

macOS / recommended path

Use cloud for this package.

What is officially true

ComfyUI does support Apple Silicon. The current official install guidance points macOS users to the latest PyTorch nightly plus the manual install path.

What we support for ALTERA2ION

For the ALTERA2ION package workflow, our supported MVP path on Mac is Comfy Cloud. That keeps setup simple without overselling local Apple hardware performance.

Best move today

If your machine is a MacBook, iMac, or iPad-based workflow companion, jump straight to Comfy Cloud. You keep the same package JSON, avoid the unsupported local hardware path, and stay aligned with the browser-first setup we can actually support for delivery.

Comfy Cloud / browser path

No GPU? No problem.

This is the cleanest setup if you do not want a dedicated local workstation. The package stays the same. You just move the compute to a browser-hosted ComfyUI environment.

Requirement What you need
Device Any Windows, Linux, or Mac laptop, plus iPad or tablet
Software Modern web browser
Local hardware No local GPU required
Network Stable internet connection
Package files Flux 2 Klein Base 9B FP8, qwen text encoder, flux2 VAE, package LoRA, supplied JSON
Comfy Cloud Flux 2 Klein 9B Official Docs

Comfy Cloud / first run

Up in a few minutes.

01 / Workspace

Create a fresh Comfy Cloud workspace for the package so you keep the model stack and JSON isolated.

02 / Model access

Accept the FLUX 2 Klein Base 9B access terms on Hugging Face first if your workspace needs that model pulled in.

03 / Files

Upload the package files your workspace needs so the LoRA and workflow sit in one clean project space.

04 / Model stack

Confirm the workspace can see the diffusion model, text encoder, VAE, and the package LoRA.

05 / Workflow

Import the supplied JSON and leave the package system prompt untouched.

06 / Queue

Drop in Image 1 and Image 2, confirm the node paths, and run the graph.

Decision guide

Local vs cloud.

Local Comfy Cloud
Setup time Longer, but full control Fastest path to first render
Hardware 22GB minimum / 24GB-class GPU No local GPU required
Best for Dedicated studio workstation Lighter devices and rapid onboarding
Devices Windows or Linux workstation Windows, Linux, Mac, iPad, or tablet
Workflow JSON Same supplied package JSON Same supplied package JSON

MVP ready

Stop switching.
Start designing.

Once ComfyUI is running, the package install is just the model stack, the LoRA, and the supplied JSON.

Browse workflows