Freepik, an online graphic design platform, has introduced a new AI image model named “F Lite,” reportedly trained exclusively on commercially licensed, safe-for-work images.
The F Lite model, developed in collaboration with AI startup Fal.ai, features approximately 10 billion parameters. It was trained with 64 Nvidia H100 GPUs over two months. F Lite joins a select group of generative AI models trained on licensed data.
According to a tweet by Javi Lopez, details reveal that the model comes in two versions: a regular version that is more predictable and prompt-faithful, and a texture version, which although more error-prone, delivers better textures.
The use of generative AI has been central in copyright lawsuits involving companies such as OpenAI and Midjourney. These models are often trained using large amounts of publicly available content, including copyrighted material, with developers frequently asserting fair use. However, creators and IP rights holders often contest this view.
Freepik offers F Lite in two variations: standard and texture. The standard version is described as more predictable, while the texture version offers superior creative compositions, despite its higher error rate.
An image generated from the standard model with the prompt “A person standing in front of a sunset, in majestic surroundings” illustrates its capability.
While Freepik does not claim F Lite surpasses leading image generators, the model is made openly available so developers can refine and enhance it. Running F Lite necessitates a GPU with at least 24GB of VRAM.
Other companies like Adobe, Bria, Getty Images, Moonvalley, and Shutterstock are also developing media-generating models based on licensed data. Depending on the legal outcomes surrounding AI copyright issues, the market for such technologies could expand significantly.
