Did Flux Krea Dev ACTUALLY Kill the Realistic Skin Problem?

Absolutely not. And after I tested myself, this statement is laughable at best. When talking about skin specifically, this model is comparable to GPT 4o.

I think all this hype around the Krea model and skin texture can be summed up into this quote by Henry Ford:

"If I asked people what they wanted, they'd say faster horses". -Henry Ford

A reshaped plastic bottle is still a plastic bottle, and while the results from Flux Krea are better, skin results are still glossy, smooth, and illogical. It's not better at high detail, production ready skin, creators are just spinning it as such for some reason. Black Forest Labs themselves don't even put this model above Kontext:

Intro To Flux Krea

First and foremost, if you haven't heard of Flux Kea Dev, this is a new model from Black Forrest labs trained on Krea which generates phenomenal images and fits right into existing Flux workflows, making it super easy to quickly load and get going! Quantized GGUF models have already been made and can be found in the resources below!

All the resources you need (workflow and models) can be found at the end of the post.

When I heard about this model, I found it from Benji's playground who is a phenomenal AI creator if you haven't seen him yet:

https://www.youtube.com/watch?v=avVCe75j0oo

Obviously, as someone actively working on AI skin, this obviously struck my attention, but after the first few seconds, I realized there is nothing to worry about.

Is AI Skin Natively Solved?

While Flux Krea Dev certainly produces images of humans with a less plastic look and slightly more imperfections/detail, it's still rather flat, simple, and too "soft". While the skin isn't shining, it's still not realistic in any capacity. And on the occasions you do get detail like freckles, pores, etc. the placement is usually illogical.

Testing Flux Krea Dev

Don't believe us? No problem - we believe that results speak for themselves too. We did some tests of our own to show Flux Krea Dev RAW outputs VS after running them through our newest Skin Refiner workflow exclusive to our platform coming soon. You tell us which you prefer!

Note with the exception of the first entry which is provided by ComfyUI's blog, all generations were created using the FP8 scaled model. It's possible that with the full model we could see better results, but I have yet to see them anywhere and the FP8 scaled model is not that terribly diluted.

The Flux Krea Dev image pictured above is the staple image on the ComfyUI blog post here:

https://blog.comfy.org/p/flux1-krea-dev-lands-on-comfyui-on

You can find all the slides shown above here: https://drive.google.com/drive/folders/1hLPR0Qp3AmdGHUDe-CX746e64Pc8NEVq?usp=sharing

Don't get us wrong whatsoever, this is a great model and the fact that this level of quality is getting more and more accessible is a great thing. If anything, Flux Krea Dev is actually a perfect model to be used with our workflows. It provides a solid foundation to build off of with a relatively low hardware cost at high resolution.

Silent Snow AI (Real Skin V1 & Upscaling Output)

And don't worry, we're testing against Enhancor too ;)

Talk very soon. I can't wait to talk about this.

Thoughts?

Resources:

Quantized Models (diffusion_models folder): https://huggingface.co/QuantStack/FLUX.1-Krea-dev-GGUF/tree/main

Full Model: https://huggingface.co/black-forest-labs/FLUX.1-Krea-dev

ComfyUI Blog: https://blog.comfy.org/p/flux1-krea-dev-lands-on-comfyui-on

Model Comparison Chart Source: https://bfl.ai/announcements/flux-1-krea-dev