F Lite joins a small group of “safe-for-work” generative models built on commercial datasets

Freepik, the popular online graphic design platform, has released a new generative AI image model that it says was trained exclusively on licensed, “safe-for-work” content. The model, called F Lite, was developed in collaboration with AI startup Fal.ai and trained over two months using 64 Nvidia H100 GPUs.

With around 10 billion parameters, F Lite is part of a growing movement toward AI models that avoid legal gray areas by using properly licensed data. Its training dataset includes approximately 80 million images from Freepik’s own commercial archive.

F Lite is available in two variations: Standard, which the company says is more prompt-accurate and predictable, and Texture, which may be less reliable but offers more creative compositions and enhanced texture rendering.

The model is designed to be open and accessible for developers to fine-tune — although Freepik acknowledges it doesn’t outperform top-tier image generators like Midjourney’s V7 or Flux from Black Forest Labs. Running F Lite also requires significant hardware: at least 24GB of VRAM on a GPU.

The release comes amid mounting legal scrutiny of generative AI companies over the use of copyrighted training data. Companies like OpenAI and Midjourney face lawsuits alleging unauthorized use of protected works. Freepik’s licensed-data approach aims to sidestep those disputes entirely.

Other companies taking a similar route include Adobe, Getty Images, Shutterstock, Bria, and Moonvalley — all part of an emerging ecosystem where data provenance could become a competitive advantage as legal standards evolve.