FLUX.2 [klein] 4B Base

FLUX.2 [klein] 4B Base

Makes and edits images from text prompts and reference pictures with an open 4B base model, built for fine-tuning, LoRA training and custom pipelines on one consumer GPU.

Made by black-forest-labs on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: FLUX.2 [klein] 4B Base is an image model from black-forest-labs. It turns a text description into a picture and can also edit a picture using one or more reference images. It has about 3.9B parameters, and this base version is undistilled, so it keeps the full training signal for people who want to adapt it.

Who it suits: Developers, researchers and artists who want an open image model they can run on one consumer graphics card and adapt with their own data.

Industry | Who uses it | How they use it | Example
Software | ML engineers | Fine-tune the base weights or train a LoRA | A custom look trained on a team's own pictures
Research | Academic labs | Study an undistilled flow model with varied outputs | Sampling many seeds to compare diversity
Design | Digital artists | Generate and edit pictures on a local machine | An edit guided by several reference images
Tooling | Pipeline builders | Wire the model into ComfyUI or diffusers workflows | A custom image pipeline with their own sampling code

Anyone who only wants the quickest possible images with no tuning can skip this base version.

Where to use it: It runs wherever you can use diffusers or ComfyUI on a suitable graphics card.

• Diffusers: the Flux2KleinPipeline class in Python
• ComfyUI, which supports the model
• The reference code and sampling scripts in the black-forest-labs GitHub repository
• Windows, macOS or Linux, for downloading the files with Python tools
• An NVIDIA RTX 3090 or 4070 class card and above, for running it

Use case | Who | What they get back
Text to image | Artists and designers | A picture made from a written description
Multi-reference editing | Designers | An edited image guided by reference pictures
Fine-tuning and LoRA training | ML engineers | A customised variant of the base weights

How to set it up: A few things need to be in place before the first run.

• Python with pip
• A graphics card with about 13GB of VRAM
• Disk space for the weights, about 7.8 GB in BF16
• A recent diffusers install, taken from its development version as the model card advises

1. Install the Hugging Face hub tools with pip.
2. Download the model files with the hf command.
3. Install or upgrade diffusers from its development version, using the pip command in the model card.
4. Save the Python script from the next section and run it.

pip install -U huggingface_hub
hf download black-forest-labs/FLUX.2-klein-base-4B

How to use it: Write a prompt, pick the image size, guidance scale, step count and a seed, then read the first image from the result. The model card's script does exactly that with Flux2KleinPipeline.

import torch
from diffusers import Flux2KleinPipeline

device = "cuda"
dtype = torch.bfloat16

pipe = Flux2KleinPipeline.from_pretrained("black-forest-labs/FLUX.2-klein-base-4B", torch_dtype=dtype)
pipe.enable_model_cpu_offload()  # save some VRAM by offloading the model to CPU

prompt = "A cat holding a sign that says hello world"
image = pipe(
    prompt=prompt,
    height=1024,
    width=1024,
    guidance_scale=4.0,
    num_inference_steps=50,
    generator=torch.Generator(device=device).manual_seed(0)
).images[0]
image.save("flux-klein.png")

Tool | What it does
Flux2KleinPipeline | The diffusers class that loads the model and makes images from a prompt
enable_model_cpu_offload | Moves parts of the model to the CPU to save VRAM
guidance_scale and num_inference_steps | Set how closely the image follows the prompt and how many sampling steps run
generator with a seed | Fixes the random seed so a run can be repeated

Example: You ask for a picture of a cat beside a sign saying hello world, at 1024 by 1024. The script returns one image and saves it as flux-klein.png. The sign text may come out distorted, since the model card warns that rendered text can be inaccurate.

For newcomers and experts: If you are new to image models, think of this as a program that makes a picture from a sentence. Start by downloading the files, then run the model card's cat script and open flux-klein.png to see what it made. You will need a recent graphics card.

Underneath, it is a rectified flow transformer trained without step or guidance distillation. That gives higher output diversity than the distilled models and leaves the weights flexible for LoRA work and fine-tuning, but it gives up the speed the family is known for. The sample run uses 50 steps, a guidance scale of 4.0 and CPU offload to fit in less VRAM.

Write a text prompt
Load Flux2KleinPipeline in diffusers
Sample with guidance and steps
Save the first image

Good to know: The card lists clear limits and banned uses, and the hardware figures come from the model page.

Title: Approximate weight size
BF16: 7.8 GB
4-bit estimate: 1.9 GB

• Limits: It cannot give factual information, and text drawn inside an image may be inaccurate or distorted.
• Bias: Output may reflect or amplify biases in the training data, and it can miss what the prompt asks for.
• Prompting: How well it follows a prompt depends heavily on the prompting style.
• License: Apache 2.0, and the card says outputs may be used commercially.
• Cost to run: none named, though you supply a GPU with about 13GB of VRAM.
• Data it reaches: model files fetched from Hugging Face, and the image file the script writes on your machine.
• Not for: breaking the law, harming minors, deceptive or harmful content, personal data used to hurt people, harassment, non-consensual intimate imagery, or fully automated high-risk decisions.

On Hugging Face, 13 November 2026

• Task: image to image
• Library: diffusers
• Parameters: 3.9B
• Weights: about 7.8 GB in BF16; about 1.9 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 743,055
• Likes: 195
• License: apache-2.0
• Languages: en
• Quantized versions: 17
• Fine-tunes: 59
• Tags: diffusers, safetensors, text-to-image, image-editing, flux, diffusion-single-file, image-to-image

Links

• Model: https://huggingface.co/black-forest-labs/FLUX.2-klein-base-4B
• Files: https://huggingface.co/black-forest-labs/FLUX.2-klein-base-4B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download black-forest-labs/FLUX.2-klein-base-4B

Then follow the usage in its model card. Nothing in this kit runs by itself.