Freepik, an online graphic design platform, has introduced a new “open” AI image model, which was announced on Tuesday. This model was developed using commercially licensed, “safe-for-work” images.
Named F Lite, the model comprises approximately 10 billion parameters, with parameters serving as the core components of the model. F Lite was created in collaboration with the AI startup Fal.ai and was trained over two months using 64 Nvidia H100 GPUs, as detailed by Freepik.
F Lite joins other generative AI models that are trained on licensed data.
In a recent statement, an individual expressed excitement about finally revealing the project, which had been under development secretly for months.
Freepik has released two versions of F Lite: standard and texture. The standard version is characterized by predictability and prompt-faithful responses, while the texture version, despite being more prone to errors, offers improved textures and creative compositions. The dataset used for training consisted of around 80 million images.
An example image generated by the standard model, based on the prompt “A person standing in front of a sunset, in majestic surroundings,” was provided by Freepik.
Freepik does not claim that F Lite produces superior images compared to leading image generators such as Midjourney’s V7 or Black Forest Labs’ Flux family. The objective was to make the model openly available for developers to customize and enhance it further.
Operating F Lite demands a GPU with at least 24GB of VRAM.
Other companies, such as Adobe, Bria, Getty Images, Moonvalley, and Shutterstock, are also developing media-generating models using licensed data. The market for these models could expand significantly, pending the outcomes of ongoing AI copyright lawsuits.
