You usually accept high latency and heavy hardware requirements, if you want great quality. You often sacrifice detail, prompt adherence, or editing flexibility if you want speed. For developers and creators building real time applications, this becomes a serious bottleneck.
Waiting several seconds per image is not acceptable for interactive tools, live previews, or production pipelines. On top of that, many high quality models are either too large, too slow, or locked behind restrictive licenses. Flux2 Klein by Black forest labs do a real time image generation and editing, without sacrificing quality.
 |
| Flux 2 Klein Image Generation showcase |
It unifies text to image, image editing, and multi reference workflows into a single compact architecture. No switching models. No hacks. Just fast, end-to-end inference that can complete in under a second.
(a) FLUX.2 klein 4B
Its the most accessible version of the family.
-4 billion parameters
-Supports Img To Img Editing and Multi-reference generation
-Designed to run on consumer GPUs like RTX 3090 or 4070
-Requires roughly 13GB VRAM
-Fully open under Apache 2.0, making it safe for commercial use.
The focus here is latency critical workflows interactive UIs, local development, and production environments where speed matters more than anything else.
 |
| Flux 2 Klein Image Editing showcase |
(b) FLUX.2 klein 9B
This is Text to Image generation the flagship small model
-9 billion parameter flow model
-Supports Single and Multi-reference generation
-Uses an 8B Qwen3 text embedder
-Step-distilled down to 4 inference steps
-Achieves quality comparable to models 5× larger, in under half a second
The trade off here is licensing. Its available under a BFL non commercial license, making it ideal for research, experimentation, and creative exploration.
Installation
1. First, install ComfyUI if you are new user. Older user need to update ComfyUI from the Manager by selecting Update ComfyUI option.
2. Download Flux 2 Klein 4B model or Flux 2 Klein 9B model from hugging face repository:-
(a) flux-2-klein-4b.safetensors - This is the distilled model that run fast but have some quality degradation.
(b) flux-2-klein-base-4b.safetensors - The base variant, run slow but have best quality.
(c) flux-2-klein-9b.safetensors- Its 9b parameter models takes almost 20GB of VRAM gives you ultimate image generation without compromising quality.
Both(a) and (b) can be run on 8GB of VRAM. Save it into ComfyUI/models/diffusion_models folder.
GGUF-
For GGUF models, make sure you have ComfyUI-GGUF custom node by City 96. If not yet done, just install from Manager by selecting Custom Nodes Manager option. Update it if already using this.
(a) Flux Klein 4b GGUF by unsloth
(a) Flux Klein 9b GGUF by unsloth
Save it into ComfyUI/models/unet folder. Text encoders and vae will same as provided below.
3. Download Text encoder from hugging face repository.
There are two models to choose from. Choose the smaller one, if you donot have enough VRAM,
(a)qwen_3_4b.safetensors - 8GB
(b)qwen_3_4b_fp4_flux2.safetensors -3.85 GB
Save it into ComfyUI/models/text_encoders folder.
4. Download VAE from hugging face repository. Save it into ComfyUI/models/vae folder.
5. Restart ComfyUI and refresh it to take effect.
Workflow
1. Download the workflows from our Hugging face repository. Then drag and drop into comfyUI.
(a) Flux2_Klein_T2Img.json (Text To Image Workflow)
(b) Flux2_Klein_Image_Edit_4b_distilled.json (Image Editing workflow)
If using the GGUF model, just load the workflow and replace the Load diffusion model loader to unet loader node.
2. Put the positive prompt into Text box and set the configuration.
KSampler Settings
Base variant-
steps- 50
CFG-4
Sampler-Euler
Distilled variant-
steps- 4
CFG-1
Sampler-Euler
The results that we are showing are not cherry picked to do the showcase the real performance of the image generation. We posted whatever we got at our first attempt.
 |
| Flux2 klein Distilled model output |
 |
| Flux 2 klein Base model output |
Prompt- This is a realistic, analog-style photo of Persian model. It captures a scene where she is attending a secret Illuminati party. She is winking, raising one hand above her head making a V-sign, and holding a highball glass in the other hand. The party features dazzling lasers and lights, and an LED screen displaying the Illuminati symbol. Aliens, Elon Musk, Donald Trump, and famous celebrities are dancing. The photo looks raw and unedited, characterized by visible film grain and the intense lighting from a flash. Cool lighting. Analog style. A candid, honest photo.
The 4B model feels like a perfect fit for developers who actually want to ship products, while the 9B model is ideal for creators who care deeply about prompt adherence and output diversity, without waiting around.
