black-forest-labs/flux-2-klein-4b

Very fast image generation and editing model. 4 steps distilled, sub-second inference for production and near real-time applications.

2.8K runs

Readme

FLUX.2 klein 4B

A compact, open-source image generation model from Black Forest Labs that brings professional image quality to smaller setups.

What is this?

FLUX.2 klein is a 4 billion parameter image model distilled from Black Forest Labs’ larger FLUX 2 base model. It’s fully open source under Apache 2.0, which means you can use it for commercial projects without licensing fees.

This model generates photorealistic images up to 4 megapixels with accurate lighting, coherent spatial relationships, and readable text rendering. It’s designed to be more powerful than other models of similar size while keeping many of the capabilities from its larger teacher model.

What makes it different?

Most small image models are trained from scratch and struggle with consistency and detail. FLUX 2 klein is different because it was distilled from a larger, more capable model. This distillation process means it inherited sophisticated understanding of lighting, materials, and composition while staying efficient enough to run on consumer hardware.

The model handles several things that typically trip up smaller models:

Text rendering — Unlike many image models that produce garbled or illegible text, FLUX.2 klein can generate clean typography in complex layouts, infographics, and user interface mockups. The text it produces is actually readable.

Spatial reasoning — The model understands how lighting works in the real world. Shadows fall where they should, reflections behave correctly, and objects maintain proper perspective relationships. This reduces the “AI look” that makes generated images feel off.

High-resolution editing — You can edit images at resolutions up to 4 megapixels while preserving detail and coherence. The model maintains geometry and texture during edits rather than hallucinating new details.

Multi-reference support — You can provide multiple reference images to maintain consistent characters, products, or styles across generations. This is useful when you need the same person to appear across multiple scenes, or when you want to generate product variations that all look like they belong to the same brand.

When to use it

FLUX.2 klein works well for workflows where you need reliable output but don’t have access to high-end hardware. It’s particularly good at:

  • Product visualization with consistent branding across variants
  • UI mockups and interface prototypes with readable text
  • Marketing materials that need exact brand colors (you can specify colors using hex codes)
  • Character consistency across multiple images for storyboarding or campaigns
  • Any situation where you need photorealistic results but can’t run larger models

The model isn’t meant to replace the larger FLUX.2 variants if you need absolute top-tier quality, but it’s surprisingly capable for its size.

How it works

FLUX.2 klein uses a latent flow matching architecture, which is a different approach from traditional diffusion models. Instead of gradually denoising an image over many steps, flow models learn direct paths between noise and clean images. This makes them more efficient and helps with consistency.

The model combines a vision-language model (based on Mistral 3) with a rectified flow transformer. The vision-language part provides semantic understanding and world knowledge, while the transformer handles spatial structure, materials, and composition. A specialized autoencoder (the FLUX.2 VAE) handles converting between pixel space and the latent space where generation happens.

Important notes

This model is fully open source under Apache 2.0. You can use it commercially, modify it, or build on top of it without needing special licenses from Black Forest Labs.

Like all FLUX.2 models, klein has been trained with safety measures to reduce the risk of generating harmful content. The training data was filtered for problematic material, and the model underwent safety fine-tuning.

The model performs best when you give it detailed, specific prompts. It can handle complex multi-part instructions and understands compositional rules better than many smaller models. You can also reference multiple images by describing them naturally in your prompt.

Getting started

You can try FLUX 2 klein on the Replicate Playground at replicate.com/playground

Model created
Model updated