Freepik's AI Image Generator Was Not Trained on Stolen Images
Freepik, a platform that provides free images, has introduced a new open-source AI image model named F Lite. Unveiled on Tuesday (yesterday), Freepik claims the model is “responsible AI” since it has been trained solely on commercially licensed, safe-for-work content.
Developed in collaboration with AI startup Fal.ai, F Lite was trained on 80 million images — a small dataset compared to other AI image generators which trained on over one billion images.
However, Freepik says that it is “potentially the largest publicly available text-to-image model trained entirely on legally sound content.” F Lite joins a small group of models trained on licensed datasets, which includes Adobe, Bria, Getty Images, Moonvalley, and Shutterstock.
“Our goal wasn’t to outshine Midjourney or Flux,” a Freepik spokesperson tells TechCrunch. “It was to make a solid, open model that developers can adapt and improve.”
⚡ IT'S FINALLY HERE!
F-Lite: our first foundational model for image generation. A collaboration between Freepik ♥️ Fal.
• Open Source• Fully commercially usable• 10B parameter DiT trained on 80M images• Trained with 100% licensed data
Link + info 🧵👇 pic.twitter.com/1k4M1qhxO9
— Javi Lopez ⛩️ (@javilopen) April 29, 2025
Generative AI is currently at the heart of multiple copyright disputes, with companies like OpenAI and Midjourney facing lawsuits for using protected content to train AI models without permission. These companies typically argue that training on such data qualifies as fair use — a legal interpretation many artists continue to challenge and one that is yet to be definitively settled in court.
F Lite is offered in two variants: standard and texture. Both were trained on Freepik’s internal library of about 80 million images. The company says the standard version is more reliable and sticks closely to prompts, while the texture variant, though more prone to mistakes, delivers richer textures and more visually creative results.
“Both models are openly licensed, with the regular and textured weights available on Hugging Face,” says Freepik. “The model’s code is also open-source, allowing you to use F Lite in ComfyUI, integrate it into your Python workflows via diffusers, or fine-tune and create custom LoRAs.”
Depending on how legal battles around AI and copyright unfold, demand for models like F Lite — trained entirely on licensed content — could see rapid growth if corporate companies are spooked by the legal uncertainty of more established image generators.