Freepik Launches F Lite AI Model Trained on Licensed Safe Images

Freepik, a prominent online graphic design platform, has launched a new open AI image model named F Lite. This model is distinct because it was trained exclusively on commercially licensed, safe-for-work images, addressing ongoing concerns about copyright infringement in generative AI development.

F Lite contains approximately 10 billion parameters and was developed in partnership with AI startup Fal.ai. The training process utilized 64 Nvidia H100 GPUs over two months, highlighting the significant computational resources invested in creating this model.

The model comes in two versions: a standard version that is more predictable and prompt-faithful, and a texture version that, while more error-prone, delivers richer textures and creative compositions. Both versions were trained on an internal dataset of around 80 million licensed images.

This approach is significant in the context of ongoing copyright lawsuits against AI companies like OpenAI and Midjourney, which often train models on massive datasets scraped from the web, including copyrighted content. Freepik’s method of using licensed data offers a transparent and ethical alternative that could influence future AI development standards.

While Freepik does not claim F Lite surpasses leading image generators like Midjourney V7 or Black Forest Labs’ Flux, the company emphasizes its goal to provide an open model that developers can tailor and improve. However, running F Lite requires a GPU with at least 24GB of VRAM, indicating substantial hardware demands.

Other companies like Adobe, Bria, Getty Images, Moonvalley, and Shutterstock are also developing generative AI models trained on licensed data, signaling a growing market trend that could expand significantly depending on the outcomes of AI copyright litigation.

The Broader Significance of Licensed-Data AI Models

The launch of F Lite represents a pivotal step toward ethical AI development by prioritizing the use of licensed content. This approach not only respects intellectual property rights but also fosters trust among creators and users. As generative AI continues to evolve, models like F Lite could set new standards for transparency and fairness in training data usage.

For developers and businesses, the availability of an open, licensed-data-driven model offers opportunities to customize AI-generated content responsibly, reducing legal risks and enhancing brand reputation. Moreover, the dual-version approach of F Lite allows for flexibility in balancing creativity and reliability based on project needs.

As AI copyright disputes unfold, the market for licensed-data AI models is poised for growth, encouraging more companies to adopt ethical training practices. This shift could lead to broader industry collaboration, innovation, and ultimately, more sustainable AI ecosystems.