Freepik, the well-known design platform, has introduced a new AI model for image generation called F Lite. The model is built entirely on licensed data, ensuring it’s safe for commercial use.
Developed in collaboration with Fal.ai, F Lite includes around 10 billion internal parameters.
The training process relied on 64 Nvidia H100 GPUs, running continuously over two months.
This approach places Freepik among a small group of companies that have chosen to build AI models exclusively on licensed datasets—at a time when major players like OpenAI and Midjourney are facing lawsuits for allegedly using copyrighted content without consent.
While many tech firms defend their practices under the “fair use” doctrine, a growing number of creators and rights holders continue to push back.
F Lite is available in two distinct versions. The Standard variant delivers accurate results that follow text prompts closely.
On the other hand, the Texture version offers more visually creative compositions with rich surface detail, though it’s slightly less consistent in precision.
Both variants are based on an internal dataset of roughly 80 million fully licensed images.
Freepik isn’t claiming that its model outperforms alternatives like Midjourney V7 or Flux by Black Forest Labs. Instead, the focus is on offering an open and modifiable model that developers can experiment with freely.
However, running F Lite locally does require powerful hardware—specifically, a GPU with at least 24 GB of VRAM.
Meanwhile, several other companies are also working on licensed-data models, including Adobe, Shutterstock, and Getty Images. This reflects a broader industry shift toward tools that respect content ownership and copyright boundaries.
The future of this field may hinge on upcoming court rulings. If the legal tide turns in favor of protecting intellectual property, licensed models like F Lite could become the new standard.
You can explore F Lite on GitHub here, or try the demo version on Hugging Face here.
