Content
summary Summary

Freepik introduces "F Lite," a new text-to-image model trained exclusively on copyright-safe material, positioning itself as a legally secure research alternative to controversial generators like Midjourney.

Ad

The company's latest generative AI model, F Lite, boasts around 10 billion parameters and was trained solely on commercially licensed images from Freepik's own collection. According to Freepik, this makes F Lite the first publicly available model at this scale to rely entirely on "safe-for-work" content.

F Lite was developed in partnership with AI startup Fal.ai. Training took place over two months on 64 Nvidia H100 GPUs, using an internal dataset of about 80 million images.

With F Lite, Freepik joins a growing list of image generators built on legally vetted data, including offerings from Adobe, Getty Images, and Shutterstock. The push for copyright-safe training sets comes as companies like OpenAI and Midjourney face a wave of lawsuits over their use of images scraped from the open internet.

Ad
Ad

Unlike some competitors, Freepik positions F Lite as an open alternative that developers can build on. The model is available on GitHub, although running it requires a GPU with at least 24 GB of VRAM. There's also a demo you can try on Hugging Face.

Human feedback fine-tunes image quality

Freepik says F Lite excels at generating illustrations and vector-style compositions, likely thanks to its training data from Freepik's own catalog.

When it comes to photorealistic images, however, the model struggles: details like skin texture or fabric are often missing, and anatomically complex scenes can result in visible errors. The model also has trouble with short prompts, as it was trained mainly on longer, descriptive text. Accurately rendering text within images remains a challenge for now.

Freepik's interpretation of the photo prompt "DSLR photograph, a red-bearded guy dancing in the rain" has a retro, early-AI feel. | Image: F Lite prompted by THE DECODER
The same prompt in Midjourney results in an artistic, photorealistic image that hardly looks AI-generated at all. | Image: Midjourney prompted by THE DECODER

The company attributes these limitations to the size of its training dataset and the amount of compute used. Freepik points to well-known scaling laws for diffusion models: quality and detail improve significantly with more data and longer training times.

Freepik's Push into Stock Media and AI Tools

Freepik is a Spain-based company specializing in licensed digital graphics for designers, businesses, and creative professionals. The platform offers a broad selection of stock media, including vector graphics, photos, illustrations, icons, and presentation templates. In addition to paid subscriptions, Freepik also provides a portion of its content for free.

Recommendation

In recent years, Freepik has expanded its investments in AI-powered tools, offering features for automatic image editing, content creation, and now, synthetic image generation.

Ad
Ad
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.
Support our independent, free-access reporting. Any contribution helps and secures our future. Support now:
Bank transfer
Summary
  • Freepik has launched F Lite, a new text-to-image model with around 10 billion parameters, trained exclusively on commercially licensed images from its own catalog and presented as a legally compliant option compared to models like Midjourney.
  • The model excels at creating illustrations and vector-style graphics, but it struggles with photorealistic outputs and short prompts, which Freepik attributes to limited data and training time.
  • F Lite is available as open source on GitHub, targeting developers interested in a legally secure foundation; using the model requires a GPU with at least 24 GB of VRAM.
Sources
Max is the managing editor of THE DECODER, bringing his background in philosophy to explore questions of consciousness and whether machines truly think or just pretend to.
Join our community
Join the DECODER community on Discord, Reddit or Twitter - we can't wait to meet you.