Freepik drops F Lite, an AI image generator trained on "copyright-safe and SFW content"

Freepik has released an open-source AI image model named F Lite. The company says F Lite is unique because it was trained exclusively on commercially licensed and safe for work content from its own internal dataset of approximately 80 million images.

The model was developed in partnership with the AI startup Fal.ai, and it has two distinct versions: a Standard Model and a Texture Model. The Standard version is intended for general-purpose image generation and is described as more predictable. The Texture Model is a specialized version that aims for richer textures and enhanced details, though it requires more detailed prompts, may be more prone to malformations, and is less effective for creating vector-style images.

F Lite"s README describes it as a 10B parameter diffusion model created by Freepik and Fal.ai, trained exclusively on copyright-safe and SFW content. Using the model locally requires a graphics card with at least 24GB of VRAM. The README also highlights that F Lite performs significantly better with detailed, longer prompts and provides workflows for ComfyUI, a popular tool for AI image generation workflows, including an advanced one that uses a SuperPrompt feature to expand simple text inputs.

This comes days after Adobe, one of Freepik"s competitors in the creative tools and stock asset market, launched a major update to its Firefly generative AI suite, introducing new image and video models and integrating third party AI models like Google Imagen 3.

The F Lite weights are available under the permissive CreativeML Open RAIL-M license, allowing developers to use and potentially fine-tune the model with their own data. Its underlying components, the T5 XXL text encoder and Flux Schnell VAE, are licensed under Apache 2.0.