Freepik, an online graphic design platform, launched a new "open" AI image model on Tuesday, and the company said it has only been trained on commercially licensed "safe working" images.

The model is called f Lite and contains about 10 billion parameters - parameters are the internal components that make up the model. According to Freepik, F lite was developed in partnership with AI startup Fal.AI and trained on 64 NVIDIA H100 GPUs over a two-month period.

F Lite joins an increasing number of generative AI models that are trained in licensing data.

We have been working on this secretly for several months! It feels great to share it in the end!

Link:

• Regular version: more predictable and timely loyal, but less art:

• Texture Version: More messy and error-prone, but offers better textures and… pic.twitter.com/gx5mipye8o

- Report Lopez ⛩️ (@javvilopen) April 29, 2025

Generative AI is a copyright litigation center for AI companies including OpenAI and Midjourney. It is often developed using a large amount of content (including copyrighted content) from public resources around the network. Most companies developing these models believe that using shields reasonably uses them to train their practices using copyrighted data without compensating the owner. Many creators and intellectual property holders disagree.

Freepik offers two flavors of standards and textures, both trained on internal datasets of approximately 80 million images. The company says the standards are more predictable and “timely faith” and the textures are more prone to errors” but provide better texture and creative work.

Here is an image from the Standard Model, which is timely "a person standing before sunset in a majestic environment."

Freepik does not claim that the F Lite produces images better than leading image generators, such as Midjourney's V7, Black Forest Labs' Flux family, or other images. The company said the goal is to publicly provide a model so developers can tailor it and improve it.

That being said, running is not an easy task. This model requires a GPU with at least 24GB of VRAM.

Other companies that develop media generation models on licensed data include Adobe, Bria, Getty Images, Moonvalley, and Shutterstock. The market could grow exponentially according to the way AI copyright litigation develops.