Over the last few years, text to image AI models have exploded in
popularity. While they have made creativity accessible to millions, one
common complaint persists the dreaded AI look. Imagine a model that not
only eliminates the plastic like AI look but also produces
photorealistic, diverse, and visually stunning images. FLUX.1 Krea Dev
developed by KreaAI, a new state-of-the-art text-to-image model that blends openness with exceptional realism.
The
model is a 12-billion parameter rectified flow transformer. Unlike most
models that fall into repetitive texture issues, this one is
deliberately trained to avoid oversaturation. Human preference tests
show that it performs on par with closed solutions like FLUX1.1 pro,
while outperforming previous open models.
Installation
1. Install ComfyUI if you have not yet. Older user need to update it from the Manager section.
2. Download Flux1 Krea Dev (flux1-krea-dev.safetensors) from official hugging face repository.
If you have low VRAM then choose the Flux Krea Dev Fp8 scaled model ( flux1-krea-dev_fp8_scaled.safetensors ) and save this into your ComfyUI/models/diffusion_models folder.
3.
These are the same models ( text encoders, vae) that already used in
basic Flux dev workflows. So downloading again is not required. But if
you want then:
Text encoders-
(a) t5xxl_fp8_e4m3fn.safetensors (For Lower VRAM)
(b) t5xxl_fp16.safetensors (Use if your VRAM is greater than 24GB)
(c) clip_l.safetensors
Save these into your ComfyUI/models/text_encoders folder.
and VAE (ae.safetensors) , save this into your ComfyUI/models/vae folder.
Workflow
1. Download the workflow (Flux.1_Krea_Dev_workflow.json) from our Hugging Face repository.
2. Drag and drop into ComfyUI.
(a) Load Flux Krea model into Load Diffusion model node.
(b) Select and upload all clip models, text encoders, and Vae into its respective node.
(c) Put prompts into prompt box
(d) Hit Run button to start generation.
KSampler Settings-
Width-1024
Height-1024
CFG-4.5
Steps-28
Here, we are trying to be as simple as it is. No, detailed prompting, for realism. Lets see how this works. The testing has been done with realism, and are not cherry picked.
Prompt: realistic mirror selfie, taken on an iPhone 14 in a well-lit modern gym, Instagram model
Prompt: selfie shot of bar girl in the new york city streets, casual in the moment
Prompt: footage of yoshi and minecraft steve fighting in a dingy alleyway, candid moment
Prompt: very pretty caucasian girl at age 18(with subtle alternative-style makeup and short, curly brown hair with soft layers and see-through side bangs), her hair is styled as a wolf cut, she is sitting in the selfie, which was taken at night in paris, with an average looking tenement visible in the background
Prompt: cctv screenshot taken from the ceiling from above the cashier, pointing towards the front of the checkout counter at a gas station. it turns out it is actually two cats in a large trenchcoat. they are buying milk and cookies.
Prompt: photo of a real spongebob working at mcdonalds in the mconalds uniform, peeking out of the drive thru window, night time
Prompt: Pikachu caught on trail cam in graveyard, grainy, weird shot
These seems very realistic with no more plastic effects.
What we observed is that to generate the best out of it, you need to be very specific detailed into your prompts. You can grab ideas from any transformers based AI tools that can do prompt enhancing.
