Freepik and the startup Fal.ai presented FLite, an open source text-to-image model with 10 billion parameters, exclusively trained with 80 million licensed images to guarantee legal security of the generated images.
FLite offers both prompt playback accuracy and creative options, requiring a robust hardware environment (24GB VRAM minimum) but providing flexibility and customization for developers and creatives.
Its launch reinforces the trend towards responsible generative AI models, based on licensed data (its creators are compensated) and promoted through open collaboration on GitHub. Let’s explore what is known about F Lite and what other alternatives exist on the market.
Origins and philosophy of FLite
Freepik and Fal.ai announced FLite on April 29, 2025, positioning it as a milestone in open and ethical AI, away from controversies over unauthorized data use.
From its conception, the project was proposed as a collaborative effort to offer a scalable and transparent AI base, not so much to compete in quality with closed solutions like Midjourney, but to serve as a foundation customizable by the community
The model has 10 billion parameters, trained for two months using 64 Nvidia H100 GPUs.
The training dataset came from 80 million fully licensed and verified images, making FLite one of the first initiatives of its scale to focus exclusively on legally clean content.
Technical requirements and limitations
For your inference (that is, to be able to run it), FLite requires a system with at least 24GB of VRAM on the GPU, which may limit its adoption among users of conventional hardware
However, this technical requirement allows the large model to be adequately handled and ensures good quality responses, although it implies higher infrastructure costs for companies and independent developers.
Two versions: standard vs. textures
The model is offered in two variants:
- Standard: Provides consistent results aligned with user indications, suitable for applications that demand accuracy.
- Textures: Prioritizes creativity, generating visual compositions of an artistic nature, although with possible minor imperfections.
Unlike other models such as Midjourney or Flux from Black Forest Labs, F Lite does not focus on achieving maximum visual quality, but rather on offering flexibility and customization.
Its open source nature, accessible through GitHub with documentation and code available, allows developers to adapt it to specific needs. This collaborative approach allows the community to propose improvements and extensions.
What other alternatives exist?
F Lite is not the only generator based on licensed data. Among the notable alternatives are:
Generative AI from Getty Images
Introduced in September 2023, it exclusively uses images licensed from Getty. It offers complete legal security, including compensation, but its payment model with undisclosed prices may limit its accessibility.
Adobe Firefly
Integrated into Creative Cloud, it uses content from Adobe Stock and licensed fonts. Its compatibility with tools like Photoshop makes it valuable for professionals, and Adobe compensates the artists involved in its training.
Shutterstock AI
Built on the Shutterstock library, it prioritizes business security. Although it provides legal guarantees, technical details about its performance are scarce.
Bria
Developed with partners like Getty, it offers customization through APIs and plug-ins. Its presence in the market, however, is less pronounced than that of other competitors.
F Lite is distinguished by its free and open source, attractive for developers and users with limited resources, although its dependence on powerful hardware can be a limitation.
Why is it important to train with licensed images?
Training AI models with properly licensed images is critical to ensuring compliance with intellectual property laws and avoiding costly copyright infringement lawsuits
By using only licensed data, companies strengthen their legal position and reduce the risk of financial sanctions, damages and injunctions that could paralyze their operations.
Conversely, relying on unauthorized content can invalidate the “fair use” defense and expose organizations to litigation based on cases such as Ross Intelligence v. Thomson Reuters or authors’ claims against Meta.
Be prepared for a more regulated future
In addition to fines, consequences can include blocking of services, withdrawal of models from the market and reputational damage that is difficult to repair.
Looking ahead, evolving regulatory frameworks such as the European Union AI Law will increasingly impose traceability, transparency and attribution requirements for training data.
Failure to comply with these regulations could lead to administrative sanctions, market limitations and even bans on the use of certain technologies in key jurisdictions.
On the other hand, respecting licenses strengthens the trust of creators and collaborators, promoting a more ethical and sustainable generative AI ecosystem, where legality and innovation coexist without compromising the creativity or value of the original authors
This post is also available in: