F Lite is a family of 10B and 7B parameter diffusion models created by Freepik and Fal, trained exclusively on copyright-safe and SFW content. The models were trained on Freepik's internal dataset comprising approximately 80 million copyright-safe images, making them the first publicly available models of this scale trained exclusively on legally compliant and SFW content.
Read our technical report for more details about the architecture and training process.
- 🎉 May, 9, 2025: F Lite 7B is released!
- April 29, 2025: F Lite is released!
| Model Version | Link | Description | Notes | HF Demo | Fal.ai Demo |
|---|---|---|---|---|---|
| Standard Model | Freepik/F-Lite | Base model suitable for general-purpose image generation | - | Demo | Fal.ai Demo |
| Texture Model | Freepik/F-Lite-Texture | Specialized version with richer textures and enhanced details | • Requires more detailed prompts • May be more prone to malformations • Less effective for vector-style imagery |
Demo | Fal.ai Demo |
| Standard 7B | Freepik/F-Lite-7B | 7B parameter version of the standard model | • Lower VRAM requirement | Demo | N/A |
F Lite can be used within ComfyUI for a more visual workflow experience. Install it into Comfy custom_nodes folder:
cd [your ComfyUI folder]/custom_nodes
git clone https://github.com/fal-ai/f-lite.git
You'll also need to install its dependencies, typically:
cd [your ComfyUI folder]/custom_nodes/f-lite
pip install -r requirements.txt
Check how to install ComfyUI plugin requirements if you have doubts.
Two example workflows are provided:
This workflow doesn't require any additional extensions and provides basic F Lite image generation functionality.
- Load the workflow by importing
F-lite-simple.jsonin ComfyUI - Perfect for quick testing without additional setup
This workflow is recommended as F Lite works significantly better with detailed, longer prompts. It uses the SuperPrompt feature to expand simple prompts into more detailed descriptions.
This workflow requires the following ComfyUI extensions:
- ComfyUI-KJNodes - Provides the
Superpromptnode for prompt expansion - ComfyUI-Custom-Scripts - Provides the
ShowTextnode for text visualization
To install these extensions:
cd [your ComfyUI folder]/custom_nodes
git clone https://github.com/pythongosssss/ComfyUI-Custom-Scripts
git clone https://github.com/kijai/ComfyUI-KJNodesAfter installing the extensions, restart ComfyUI and load the F-lite-superprompt.json workflow.
pip install -r requirements.txtYou can generate images using the provided generate.py script:
python -m f_lite.generate \
--prompt "A photorealistic landscape of a mountain lake at sunset with reflections in the water" \
--output_file "generated_image.png" \
--model "Freepik/F-Lite" \
--width 1344 \
--height 896 \
--steps 30 \
--guidance_scale 6 \
--seed 42Here's a basic example of how to use the F Lite pipeline:
import torch
from f_lite import FLitePipeline
# Trick required because it is not a native diffusers model
from diffusers.pipelines.pipeline_loading_utils import LOADABLE_CLASSES, ALL_IMPORTABLE_CLASSES
LOADABLE_CLASSES["f_lite"] = LOADABLE_CLASSES["f_lite.model"] = {"DiT": ["save_pretrained", "from_pretrained"]}
ALL_IMPORTABLE_CLASSES["DiT"] = ["save_pretrained", "from_pretrained"]
pipeline = FLitePipeline.from_pretrained("Freepik/F-Lite", torch_dtype=torch.bfloat16)
pipeline.enable_model_cpu_offload() # For less memory consumption. Alternatively, pipeline.to("cuda")
pipe.vae.enable_slicing()
pipe.vae.enable_tiling()
# Generate an image
output = pipeline(
prompt="A photorealistic 3D render of a charming, mischievous young boy, approximately eight years old, possessing the endearingly unusual features of long, floppy donkey ears that droop playfully over his shoulders and a surprisingly small, pink pig nose that twitches slightly. His eyes, a sparkling, intelligent hazel, are wide with a hint of playful mischief, framed by slightly unruly, sandy-brown hair that falls in tousled waves across his forehead.",
height=1024,
width=1024,
num_inference_steps=30,
guidance_scale=3.0,
negative_prompt=None,
)
# Save the generated image
output.images[0].save("image.png")The model requires a card with at least 24GB of VRAM to operate, although using quantization has not been explored and it is expected to reduce the memory footprint even more.
You can enable Adaptive Projected Guidance by setting the parameter apg_config to APGConfig(enabled=True).
Using negative prompts can help improve the quality of the generated images.
It is possible to fine-tune F Lite using your own data. Read the Fine-tuning documentation for more information.
f-lite includes a Gradio-based GUI that provides an intuitive interface for image generation.
python f_lite_gradio_gui.py
- Basic parameter configuration
- Augmented Parallel Guidance for more consistent results
- Preset resolutions for common aspect ratios (square, portrait, landscape)
- Continuous generation mode with random seed iteration
- Dynamic prompt enhancement with wildcard support (
{option1|option2|option3}syntax, as well as__filename__for .txt-file wildcards) - Enhanced prompts using SuperPrompt
- Controls for randomization with SuperPrompt and wildcards while using fixed seeds
- Prompt prefix / suffix
This model uses T5 XXLand Flux Schnell VAE
The F Lite weights are licensed under the permissive CreativeML Open RAIL-M license. The T5 XXL and Flux Schnell VAE are licensed under Apache 2.0.
If you find our work helpful, please cite it!
@article{ryu2025flite,
title={F Lite Technical Report},
author={Ryu, Simo and Pengqi, Lu and Mart\'in Juan, Javier and de Prado Alonso, Iv\'an},
year={2025}
}