Flux Krea (FLUX.1-Krea-dev) is a text-to-image model from Black Forest Labs fine-tuned by Krea AI. It produces highly realistic and detail-rich images and is compatible with ComfyUI workflows.

What models do I need for Flux Krea in ComfyUI?

You need the flux1-krea-dev_fp8_scaled.safetensors diffusion model, clip_l.safetensors and t5xxl_fp8_e4m3fn_scaled.safetensors text encoders, and ae.safetensors as the VAE.

Where do I place Flux Krea model files in ComfyUI?

Diffusion models go in ComfyUI/models/diffusion_models/, text encoders in ComfyUI/models/text_encoders/, and the VAE in ComfyUI/models/vae/.

How much VRAM do I need for Flux Krea?

The fp8 scaled version of the diffusion model significantly reduces VRAM requirements. 8–12 GB of VRAM is recommended; 16 GB or more gives comfortable headroom for higher resolutions.

How to Install Flux Krea in ComfyUI — Step-by-Step Guide (2026)

Q: Is Flux Krea free to use locally?

Yes. Flux Krea model weights are available on Hugging Face and can be run locally in ComfyUI at no cost, subject to the FLUX.1-dev license terms.

Flux-Krea Text-to-Image Workflow

What you'll need: An updated ComfyUI installation, the flux-krea JSON workflow file, and three model files (diffusion model, text encoders, VAE) downloaded from Hugging Face.

Creating images from text is easier than ever with the Flux-Krea workflow for ComfyUI. Flux Krea (FLUX.1-Krea-dev) is a powerful text-to-image model that produces highly realistic, detail-rich results directly from natural language prompts. This guide walks you through every step — from downloading the workflow to running your first generation.

The setup takes about 10–15 minutes. Once the models are in place, generating images is as simple as typing a prompt and pressing run.

Step 1 — Download the Workflow File

First, make sure your ComfyUI is updated to the latest version. Then download the JSON workflow below and drag it directly into your ComfyUI canvas to load it.

Download flux-krea JSON Workflow

Drag the downloaded flux-krea.json file into your ComfyUI browser window to load the workflow automatically.

💡 Tip: If ComfyUI shows missing node errors after loading the workflow, go to ComfyUI Manager → Update All to ensure all custom nodes are current, then reload the workflow.

Step 2 — Download Required Models

Flux Krea requires three types of model files. Download each from Hugging Face and place them in the correct ComfyUI subfolders as shown in the folder structure below.

Diffusion Model

Choose one of the following — the fp8 scaled version is recommended for lower VRAM usage:

flux1-krea-dev_fp8_scaled.safetensors (recommended — lower VRAM)
flux1-krea-dev.safetensors (full precision — more VRAM required)

Text Encoders

You need both of the following text encoder files:

clip_l.safetensors
t5xxl_fp8_e4m3fn_scaled.safetensors (fp8 — recommended) or t5xxl_fp16.safetensors (fp16 — higher quality, more VRAM)

VAE

ae.safetensors

Required Folder Structure

Place each file in the correct subfolder. The filenames must match exactly:

📂 ComfyUI/
└── 📂 models/
    ├── 📂 diffusion_models/
    │   └── flux1-krea-dev_fp8_scaled.safetensors
    │       (or flux1-krea-dev.safetensors)
    ├── 📂 text_encoders/
    │   ├── clip_l.safetensors
    │   └── t5xxl_fp8_e4m3fn_scaled.safetensors
    │       (or t5xxl_fp16.safetensors)
    └── 📂 vae/
        └── ae.safetensors

Step 3 — Workflow Setup

Once the workflow is loaded and all model files are in place, configure each node to point to the correct file. In each node listed below, open the dropdown and select the matching filename:

In theLoad Diffusion Modelnode, select:flux1-krea-dev_fp8_scaled.safetensors

In theLoad CLIPnode, select:clip_l.safetensors

In theLoad text_encodersnode, select:t5xxl_fp8_e4m3fn_scaled.safetensors

In theLoad VAEnode, select:ae.safetensors

Next, enter your prompt. Here's a ready-to-use example to get started:

Starter prompt

A realistic young woman standing on the cracked platform of an old abandoned train station at sunset. Rusty steel beams, chipped paint, and overgrown grass between the tracks are rendered in high deta…

💡 Tip: Adjust width and height to match your target aspect ratio. You can also tune CFG and Steps based on your GPU — higher steps improve quality but take longer.

Run the workflow

Click the Queue Prompt button, or press Ctrl + Enter (Windows / Linux) or Cmd + Enter (Mac) to start generating your image.

Step 4 — Example Generations

Here are three example prompts with their outputs generated using Flux Krea in ComfyUI. Each prompt is ready to copy and use directly in your workflow.

Example 1 — Jungle beach scene, cinematic atmosphere

Prompt used:

At the lush boundary where a dense, humid jungle meets a sun-bleached beach, a beautiful woman seeks refuge from the heat. She wears a skimpy, leopard-print bikini that blends with the wild surroundin…

Output:

Example 2 — Cozy cat portrait, soft window lighting

Prompt used:

(photorealistic:1.4), (masterpiece:1.3), (intricate details:1.2), A charming orange and white domestic shorthair cat, with soft and fluffy fur, lying on a plush and cozy bed, gazing out of the window …

Output:

Example 3 — Muscular Doberman in fashion sunglasses, studio lighting

Prompt used:

(best quality, 8K, high resolution, masterpiece), ultra-detailed, hyper-realistic photography, cinematic lighting — a majestic, intensely muscular Doberman stands proudly against a deep matte black ba…

Output:

Note: In most cases, results closely match the prompt — realism and detail are impressive with Flux Krea. Your output may vary slightly depending on prompt clarity, CFG value, and step count.

Final Thoughts

Flux Krea is one of the most capable text-to-image models available for local use in ComfyUI. Once the models are in place, the workflow is straightforward — and the results speak for themselves across portraits, product shots, landscapes, and more.

Setup is simple once model files are in the correct folders.
Works best with clear, detailed, and well-structured prompts.
The fp8 scaled model significantly reduces VRAM requirements without a major quality drop.
Experiment with CFG and step count to balance speed and quality for your GPU.

Ready to go? Download the workflow above, place your models, and start generating. If you run into issues, the ComfyUI Troubleshooting Guide covers the most common errors and fixes.

Frequently Asked Questions

Flux Krea (FLUX.1-Krea-dev) is a text-to-image model built on the FLUX.1-dev architecture and fine-tuned by Krea AI. It produces highly realistic and detail-rich images from natural language prompts and runs locally in ComfyUI.

You need three types of files: the diffusion model (flux1-krea-dev_fp8_scaled.safetensors), two text encoders (clip_l.safetensors and t5xxl_fp8_e4m3fn_scaled.safetensors), and the VAE (ae.safetensors). All are available on Hugging Face.

Diffusion models go in ComfyUI/models/diffusion_models/, text encoders go in ComfyUI/models/text_encoders/, and the VAE goes in ComfyUI/models/vae/. Filenames must match exactly what appears in the node dropdowns.

Yes. The Flux Krea model weights are publicly available on Hugging Face and can be run locally in ComfyUI at no cost, subject to the FLUX.1-dev license terms. Check the Hugging Face model page for the full license.

The fp8 scaled diffusion model reduces VRAM requirements significantly. 8–12 GB of VRAM is workable at standard resolutions; 16 GB or more gives comfortable headroom for larger outputs. Lower resolutions require less VRAM.