1,170,445 views 19 Aug 2024. Take a first look at a groundbreaking new open-source AI image generation model Flux, which rivals Stable Diffusion and MidJourney. Learn how to run “flux.1” locally and train a LoRA with your own custom data.
TRANSCRIPT (auto-gen)
Last week two brand new AI image generators dropped imagin 3 from Google and grock 2 from Elon both can generate hyperrealistic images with accurate text but only one of them is open source and uncensored can you guess which one that’s right the answer is neither the real hero is actually flux from Black Forest Labs a new image gen model that’s taking the World by storm and is the underlying model that powers grock’s images people are calling it The Mid Journey killer and nextg stable diffusion replacement a model so realistic and dangerous that it must be stopped Google Deep Mind even released a paper this month that studied the ways people abuse generative Ai and shockingly intimate images wasn’t the top result apparently the real danger is impersonation which is something that flux happens to be really good at in today’s video you’ll learn how to run this dangerous new AI locally fine-tune it with your own custom data and turn the results into video is so you can level up your AI girlfriend in ways you never thought possible it is August 19th 2024 and you’re watching the code report people on the internet are outraged by this photo which was generated by roock and its NAB or flux although one could argue it’s wholesome and shows how libtards and cuckservatives can work together to make the world a better place it’s obviously fake but look at this photo it’s benign and realistic but was generated with flux is so realistic that it doesn’t even feel uncanny in just a minute we’ll learn how to use flux and fine-tune it on our own images but Google also just released its own new image generator called image gen 3 it has an amazing UI that also uses AI to generate prompts as you write them and the image quality is actually quite impressive it’s it’s a huge improvement from image Gen 2 which had to be taken down for generating ethnically diverse Nazis which was so bad that even Sergey said quote we definitely messed up imin 3 is far better tuned for generating Nazis but it definitely won’t let you impersonate people or generate anything even slightly offensive but it does have some cool features like an editor where you can regenerate different parts of the image if you’re serious about doing crazy stuff with AI though the hottest model in town right now is flux which was created by former employees of stability AI who worked on stable diffusion that’s pretty exciting because stable diffusion 3 medium was fairly disappointing and now a new bar is set for open source models that will likely only be rivaled by stable diffusion 3 large in addition there’s already a bunch of lauras or low rank adaptations that find toune flux for special use cases which you can find on places like civid AI in fact you can even find tune it with pictures of yourself like Matt wolf did to generate images for your Instagram to make it look like your life is actually cool the easiest way to Tri flux is to use a hosted version on a website like foul which not only has the base model but also lauras like flux realism which can generate photo realistic results consistently but one thing to know is that there are three different flux models flux Pro flux Dev and flux Schnell the smallest model Schnell is the only one that’s licensed under Apache 2.0 if you plan on using it commercially that’s the model you’d want to go with otherwise you can access the pro model through the Black Forest Labs API but when it comes to experimentation the dev model will give you the highest quality and efficiency it just can’t be used commercially let’s go ahead and run it loc Al now one option is to use the hugging face diffusers library in a python script like this which will automatically download the model allowing you to prompt it and run it on your GPU with just a few lines of python but if you’re not well endowed with a big GPU make sure to enable the CPU offload mode from there assuming you have a virtual environment with your dependencies installed you can run the script and start generating images locally that’s a good start but the real power of open weights is being able to fine-tune your own models there’s a few different open source projects that make that way easier like simple tuner or xlux if you want to use a node-based workflow they have a setup for comfy UI or some training scripts that you can customize in yaml to train your own Laura you’ll first need some data like a folder of images with corresponding Json files that contain a caption for that image like if you’re a deranged stalker that could be some images of your ex-girlfriend or if you’ve never had a girlfriend maybe a celebrity you’re obsessed with the important concept to remember is garbage in garbage out you’ll get the best results with high quality data from there you can fine- tune your hyperparameters then start training your own model with a single command it’s just that easy pretty crazy how easy that is but the real benefit of this technology is that it allows us to generate our own AI girlfriends and boyfriends here’s what a full stack AI partner looks like step one build a data set of about 20 images and captions of what you want your partner to look like the step two to train aura based on flux you now have a model that can generate realistic Unique Images is step three give your partner a voice using a tool like 11 Labs which can even be cloned from a real human then finally step four use a tool like pabs to generate a video and lip sync mode to match the voice to the video congratulations you just cured loneliness don’t hate me because I’m artificial this has been the code report thanks for watching and I will see you in the next one.