Freepik’s F Lite: An Open AI Image Model Built on Licensed Data
▼ Summary
– Freepik has launched F Lite, an AI image generation model trained exclusively on licensed, commercially safe images to address copyright issues in generative AI.
– Unlike many AI models that use internet-scraped datasets, F Lite is built on Freepik’s internal library, aiming to avoid legal controversies over copyright.
– F Lite is a 10-billion-parameter model trained on 80 million images using 64 Nvidia H100 GPUs, offering two versions: ‘standard’ for predictability and ‘texture’ for creative detail.
– The model is openly available under a CreativeML Open RAIL-M license, though it requires a high-performance GPU with at least 24GB of VRAM.
– Freepik’s approach mirrors a trend among companies like Adobe and Getty Images, prioritizing licensed data to mitigate legal risks in AI development.
Freepik Charts a Course Through AI Copyright Waters
Graphic design resource Freepik has officially released F Lite, a new AI image generation model notable for its training foundation: exclusively licensed, commercially safe images. This move places Freepik among a cohort of companies directly addressing the copyright controversies swirling around generative AI development.
The standard practice for training many large-scale AI models involves processing vast datasets scraped from the internet, often including copyrighted works without explicit permission, leading to high-profile lawsuits against firms like OpenAI and Midjourney. While these companies often argue fair use, Freepik, partnering with AI startup Fal.ai, is deliberately building its model on a less contentious base – its own internal library.
Under the Hood: Training F Lite
According to Freepik and Fal.ai, F Lite is a 10-billion-parameter model. Its training utilized roughly 80 million “copyright-safe” and “safe-for-work” images drawn entirely from Freepik’s internal collection over two months, powered by a cluster of 64 Nvidia H100 GPUs. This licensed dataset approach aims to provide users, particularly commercial ones, with greater peace of mind regarding image provenance.
Freepik offers the model in two versions: ‘standard’, designed for predictability and closely following text prompts, and ‘texture’, which may be less predictable but aims for enhanced textural detail and creative flair. Both versions are available.
Open Access, Significant Requirements
Freepik has made F Lite openly available, distributing it via platforms like Hugging Face under a CreativeML Open RAIL-M license. The company isn’t necessarily claiming F Lite surpasses market leaders in sheer image quality but emphasizes its open nature, allowing developers to adapt and build upon a model trained on ethically sourced data.
However, potential users should note the hardware demands. Running F Lite requires a capable GPU equipped with at least 24GB of VRAM, placing it beyond the reach of typical consumer-grade computers and targeting users with more specialized setups.
The Growing Licensed-Data Trend
Freepik’s strategy aligns with other significant players like Adobe, Getty Images, and Shutterstock, who are also developing generative models using licensed content libraries. As legal challenges surrounding AI training data continue, models built with clearly documented, rights-cleared data could become increasingly valuable, particularly for commercial applications where copyright indemnity is a concern. Freepik’s F Lite adds another option for developers seeking such reassurances.
