Generative AI image generation : FLUX an amazing open source model

[podumet]

Hi

A few month ago, FLUX.1 models series has been released. I give them a try and was extremely impressed by the results, for me it's the best open source image generation model available now.
More here : https://flux-1.ai/

Vanilla off-the-shelf FLUX models are great and address some of the main issues with existing ones, like Stable Diffusion (which struggles with rendering hands correctly and writing text).

But the great thing is that FLUX is open source, meaning it's possible to fine-tune the model with a consumer-grade GPU (a 24GB VRAM GPU, which costs around $2000, can handle it). Since I have both the hardware and experience in fine-tuning such large models, I gave it a try, and the results are extremely impressive.

I succeeded in adding characters to the model with remarkable results.

Example:

I trained the model with just a few pictures of a French actress (10 in total), the real actress looks like this:

The training last for about 1 hour.

Now the model knows who is Laure, and I can use her in the prompt.

First prompt : "Laure, sitting in the grass in Paris in a public park with a yellow dress and red shoes, Eiffel Tower back ground, natural lighting of sunset"​

Second test more tricky with text

Prompt : "Laure floating in the air in the International Space Station with a blue suit holding with 2 hands a paper with the text "Hello ACF"​

You will note the quality of the hands

Prompt "Laure in manga anime drawing with a short black dress"​

And a more tricky one with an indirection on the prompt

prompt : "A man is standing, viewed from the back, painting a portrait of Laure on an watercolor paint board. He is holding a paintbrush in one hand and a palette in the other"​

Here we are with 10 photos of someone, you can create virtual photo shoot (for the best) or terrible deep fake (for the worst).

We have 10 photos of someone, and with them, you can create either a virtual photoshoot (at best) or a terrible deep fake (at worst).

So, what do you think? Will it be the worst or the best?

 

[Nigerian Prince]

I'm curious, what was your thought process when posting this

 

[podumet]

No special thoughts, just wanted to mention that this exists, is accessible for free, and is becoming extremely powerful.

By the way, I believe these models can be both useful and threatening for original content creators and artists and generate some debate here, no more, ,no less

....
Alright, I admit that I was also quite happy with the generated images and wanted to show off a bit.

 

[Nigerian Prince]

No special thoughts, just wanted to mention that this exists, is accessible for free, and is becoming extremely powerful.

By the way, I believe these models can be both useful and threatening for original content creators and artists and generate some debate here, no more, ,no less

....
Alright, I admit that I was also quite happy with the generated images and wanted to show off a bit.

Ok well here's something to think about. You have this software available that allows you to feed it images of anyone and generate realistic-ish images of them doing whatever you want wherever you want, regardless of what the actual person the images are based on thinks of them

 

[podumet]

Yes, thank you for your point—and for starting the debate.

But the thing is, I'm not the only one experimenting with this technology. I'm just an individual playing around with it.

Just imagine what the industry will do with it... People stealing your content today might steal your entire image tomorrow.

Right now it's pictures, but the first solutions for generating videos are emerging (like DeepBrain) and will also be freely available to anyone very soon.

This is the bad side.

But there's also a good side: you can create your own digital twin by yourself, and this is the benefit of having this open model not only in the hands of a select few. By doing so, you're able to create new ways to engage with your audience. It's about being proactive and finding ways to adapt...

 

[MarieElise]

Yes, thank you for your point—and for starting the debate.

But the thing is, I'm not the only one experimenting with this technology. I'm just an individual playing around with it.

Just imagine what the industry will do with it... People stealing your content today might steal your entire image tomorrow.

Right now it's pictures, but the first solutions for generating videos are emerging (like DeepBrain) and will also be freely available to anyone very soon.

This is the bad side.

But there's also a good side: you can create your own digital twin by yourself, and this is the benefit of having this open model not only in the hands of a select few. By doing so, you're able to create new ways to engage with your audience. It's about being proactive and finding ways to adapt...

I noticed in your first post here you say you might be needing the support of a small community soon. Tell us more about that upfront.

Along with more specifics regarding the AI project you are working on which analyzes live cam models streams. This is the most relevant info I would like to know, regarding your presence here, before I consider engagement in any topics started by you. I'm sure I'm not the only one. Specifics plz.