The latest batch of AI-generated photos to take social media by storm features close-ups of men and women talking into microphones presumably onstage at what appears to be a conference or some kind of event. Having first surfaced on a Reddit thread dedicated to AI-generated content, the realistic depiction of the skin, hair, and wrinkles of the men and women in the images have sparked a lot of buzz online. These pictures were generated using a newly launched text-to-image AI model called FLUX.1 that has been developed by Black Forest Labs, a company based in Germany. These TEDx speakers aren't real. They're made with the new Flux realism LoRA. Flux is founded by the core engineers of stable diffusion. It’s crazy good. Here's the Lora file: Here's the ComfuUI workflow: FluxRealLoraWorkflowjson.json… pic.twitter.com/n6kQwVPDTQ — Grant (@GrantAIAgent) August 9, 2024 While images of this kind have surface online in the past, the ones created using FLUX.1 demonstrate that such AI models are now capable of showing fake people with nothing evidently off about them (except for the text on their badges). Hence, they serve as an indication that the technology is not only becoming increasingly sophisticated but also more accessible since FLUX.1 is an open-source AI model. However, it must be noted that the AI-generated photos created using FLUX.1 that have gone viral were further enhanced using a Low-Rank Adaptation (LoRA) technique that is meant to refine and optimise the outputs of large language models. What is FLUX.1 and who are its makers? Launched on August 1, FLUX.1 is available for users in three different versions. The ‘Pro’ version of the AI model is the high-end one designed for commercial use. Meanwhile, the ‘Dev’ version has open weights and built for non-commercial use, while the faster version of FLUX.1 with open weights is called ‘Schnell’ which translates to quick or fast in German. The architecture of the AI model reportedly employs a combination of transformer and diffusion techniques with a parameter count of 12 billion. For context, the suite of Stability Diffusion 3 models have parameter size ranging from 800 million to 8 billion. FLUX.1 has been further optimised through training methods such as flow matching, Black Forest Labs said in a blog post. However, company did not identify the source of the data used to train its AI model. FLUX.1 has been built by Black Forest Labs, an AI startup that was founded by a group of AI researchers and engineers, some of whom previously held key positions at Stability AI . In fact, Robin Rombach, Andreas Blattmann, and Dominik Lorenz were reportedly working at the AI company up till the launch of its Stability Diffusion 3 (SD3) model. Recently, Black Forest Labs managed to raise funds worth over $31 million in a seed funding round that saw investments from Silicon Valley investors such as Andreessen Horowitz as well as General Catalyst and MätchVC. AI researcher Matthias Bethge and former Disney president Michael Ovitz have also been brought on as advisers to Black Forest Labs, according to a report by ArsTechnica. What are the capabilities of FLUX.1? FLUX.1 seems to be capable of generating human hands and legs. The accurate rendering of human limbs was a problem in images generated using AI models due to inadequacies in the training datasets. The company behind FLUX.1 announced a new video model. One thing that stands out is the consistency of the character, clothes and style between the cuts at the beginning of the video. This looks extremely promising. pic.twitter.com/hsLCrtAdcW — Halim Alrasihi (@HalimAlrasihi) August 6, 2024 “All FLUX.1 model variants support a diverse range of aspect ratios and resolutions in 0.1 and 2.0 megapixels,” the company said. Moving forward, Black Forest Labs said that it is working on a text-to-video generator rivalling OpenAI's Sora, Runway's Gen-3 Alpha, and Kuaishou's Kling. "Our video models will unlock precise creation and editing at high definition and unprecedented speed," the post read. The weights of FLUX.1 [dev] and FLUX.1 [schnell] are openly available on AI developer platforms like Hugging Face. For direct use, these models can be accessed through AI cloud-hosting platforms Replicate and Fal. Stock image website Freepik has also integrated FLUX.1 as part of its AI toolbox.
