AB
AB
AiBoss

All Categories

project

F-Lite - A text-based graph model open-sourced by Freepik and FAL.

F-Lite is a 10-parameter text-based image model open-sourced by the Freepik team in collaboration with FAL. It is trained on Freepik's proprietary 80M dataset and supports commercial use. F-Lite uses T5-XXL as the text encoder, based on extracting the 17th...

What is F-Lite?

F-Lite is a 10-parameter text-to-image model open-sourced by the Freepik team in collaboration with FAL. Trained on Freepik's proprietary 80M dataset, it supports commercial use. F-Lite uses T5-XXL as the text encoder, injecting extracted features from the 17th layer into the DiT model. Training includes pre-training at 256 and 512 resolutions, and post-training at 1024 resolution, resulting in high training costs. A special version, F-Lite Texture, optimized for rich textures and detailed cues, has been released.

F-Lite's main functions

  • Text to Image GenerationThe user inputs a text description, and the model generates an image that matches the description.
  • Commercial LicenseThe model was trained on the copyright-safe dataset provided by Freepik, and the generated images can be used for commercial purposes.
  • Multi-resolution trainingIt supports image generation at resolutions of 256, 512, and 1024 to meet the needs of different scenarios.
  • Special version optimizationIntroducing the F-Lite Texture version, optimized for rich textures and detailed hints.

F-Lite's technical principles

  • Diffusion Model ArchitectureThis method utilizes a reverse diffusion process to gradually transform random noise into meaningful images. A text-based conditional diffusion model, combined with a text encoder, injects textual features into the image generation process.
  • Text encoderThis method uses T5-XXL as a text encoder to extract text features. Features are extracted from the 17th layer of the T5-XXL encoder, rather than the last layer, to better capture the semantic information of the text. The text features are injected into the diffusion model using a cross-attention mechanism to ensure that the generated image is highly relevant to the text description.
  • Training strategy:
    • Multi-resolution pre-trainingThe model is pre-trained at 256 and 512 resolutions to learn the basic features of the image.
    • High-resolution post-trainingPost-training at 1024 resolution generates higher quality images.
    • Reinforcement learning training: Reinforcement learning training based on GRPO (Gradient-based Reinforcement Policy Optimization) improves the diversity and quality of generated images.
  • Optimization technologyLearnable register tokens are introduced to better align text and image features. Residual connections improve the model's training stability and efficiency. μ-Parameterization optimizes the diffusion process, enhancing the quality of generated images.

F-Lite project address

Application scenarios of F-Lite

  • Creative DesignIt provides inspiration and visual materials for advertising, posters, illustrations, and other designs, improving design efficiency and creative diversity.
  • Content creationGenerate images for social media, blogs, etc., to enrich the visual effects of content and enhance its appeal and dissemination.
  • Game developmentIt can quickly generate game characters, scenes, and complex textures, accelerating the game design and development process.
  • Education and LearningThe system generates relevant images based on the teaching content, helping students to better understand and remember the material, thus improving learning outcomes.
  • Business and EnterpriseGenerate product display images, brand promotional images, etc., for business promotion and brand building, to enhance brand image and market competitiveness.
Back