FLUX.2 [klein] 4B & 9B - Fast local image editing and generation

High-speed image editing with compact Flux 2 Klein models, now supported in ComfyUI.

FLUX.2 [klein] 4B & 9B are the fastest image models in the Flux family, unifying image generation and image editing in a single, compact architecture.

Designed for interactive workflows, immediate previews, and latency-critical applications, FLUX.2 [klein] delivers state-of-the-art image quality with end-to-end inference around one second on distilled variants—enabling creative iteration at a pace that wasn’t previously practical with diffusion models.

---

Two Models, Two Types

FLUX.2 [klein] is released across two model types, each available at 4B and 9B parameters:

Base (Undistilled)

• Full training signal and model capacity

• Optimized for fine-tuning, LoRA training, and post-training workflows

• Maximum flexibility and control for research and customization

Distilled (4-Step)

• 4-step distilled for the fastest inference

• Built for production deployments, interactive applications, and real-time previews

• Optimized for speed with minimal quality loss

---

Model Lineup and Performance

9B distilled — 4 steps · ~2s (5090) · 19.6GB VRAM

9B base — 50 steps · ~35s (5090) · 21.7GB VRAM

4B distilled — 4 steps · ~1.2s (5090) · 8.4GB VRAM

4B base — 50 steps · ~17s (5090) · 9.2GB VRAM

Both sizes support text-to-image and image editing, including single-reference and multi-reference workflows.

---

HuggingFace Repositories - BF16

https://huggingface.co/black-forest-labs/FLUX.2-klein-9B

https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9B

https://huggingface.co/black-forest-labs/FLUX.2-klein-4B

https://huggingface.co/black-forest-labs/FLUX.2-klein-base-4B

---

HuggingFace Repositories - FP8

https://huggingface.co/black-forest-labs/FLUX.2-klein-9b-fp8

https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9b-fp8

https://huggingface.co/black-forest-labs/FLUX.2-klein-4b-fp8

https://huggingface.co/black-forest-labs/FLUX.2-klein-base-4b-fp8

---

HuggingFace Repositories - NVFP4

https://huggingface.co/black-forest-labs/FLUX.2-klein-9b-nvfp4

https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9b-nvfp4

https://huggingface.co/black-forest-labs/FLUX.2-klein-4b-nvfp4

https://huggingface.co/black-forest-labs/FLUX.2-klein-base-4b-nvfp4

---

9B vs 4B: Choosing the Right Scale

FLUX.2 [klein] 9B Base

The undistilled foundation model of the Klein family.

• Maximum flexibility for creative exploration and research

• Best suited for fine-tuning and custom pipelines

• Ideal where full model capacity and control are required

FLUX.2 [klein] 9B (Distilled)

A 4-step distilled model delivering outstanding quality at sub-second speed.

• Optimized for very low-latency inference

• Near real-time image generation and editing

• Available exclusively through the Black Forest Labs API

---

FLUX.2 [klein] 4B Base

A compact undistilled model with an exceptional quality-to-size ratio.

• Efficient local deployment

• Strong candidate for fine-tuning on limited hardware

• Flexible generation and editing workflows with low VRAM requirements

FLUX.2 [klein] 4B (Distilled)

The fastest variant in the Klein family.

• Near real-time image generation and editing

• Built for interactive applications and live previews

• Sub-second inference with minimal overhead

---

Editing Capabilities

Both FLUX.2 [klein] 4B models support image editing workflows, including:

• Style transformation

• Semantic changes

• Object replacement and removal

• Multi-reference composition

• Iterative edits across multiple passes

Single-reference and multi-reference inputs are supported, enabling controlled transformations while maintaining visual coherence.

Use Image Edit to explore multiple angles of a single subject

---

Use multiple input images to precisely guide generation

---

Iterate on color and material texture for precise control

---

Get Started

1. Update to the latest version of ComfyUI

2. Browse Templates and look for Flux.2 Klein 4B & 9B under Images, or download the workflows

3. Download the models when prompted

4. Upload your image and adjust the edit prompt, then hit run!