F Lite is a 10 billion parameter diffusion model designed for high-quality, copyright-safe image generation. It is trained exclusively on SFW content from Freepik's dataset, making it suitable for users prioritizing legal compliance and ethical AI development.

How It Works

F Lite is built upon a Diffusion Transformer (DiT) architecture, leveraging a large-scale dataset of 80 million curated images. This approach allows for detailed image synthesis with a focus on rich textures and enhanced visual fidelity, distinguishing it from models trained on broader, less controlled datasets.

Quick Start & Requirements

• Installation: pip install -r requirements.txt
• Prerequisites: Requires a GPU with at least 24GB of VRAM.
• Usage: Command-line generation via python -m f_lite.generate or integration with the diffusers library.
• ComfyUI: Supports ComfyUI workflows, including an advanced setup with prompt expansion via ComfyUI-KJNodes and ComfyUI-Custom-Scripts.
• Resources: Model weights are available on Hugging Face.

Highlighted Details

• 10B parameter diffusion model.
• Trained exclusively on 80 million copyright-safe, SFW images.
• Offers a standard and a texture-specialized model version.
• Supports prompt expansion for improved results.

Maintenance & Community

• Developed by Freepik and Fal.
• Technical report available for detailed architecture and training insights.
• Project updates are announced, with demos provided on Hugging Face and Fal.ai.

Licensing & Compatibility

• F Lite weights are licensed under the CreativeML Open RAIL-M license.
• T5 XXL and Flux Schnell VAE components are licensed under Apache 2.0.
• Permissive licensing allows for commercial use and integration into closed-source projects.

Limitations & Caveats

The texture model may require more detailed prompts and can be less effective for vector-style imagery. The model requires significant VRAM (24GB+), though quantization has not been explored to reduce this footprint.