import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("Freepik/F-Lite-7B", dtype=torch.bfloat16, device_map="cuda")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]F Lite 7B Model Card
F Lite 7B is a 7 B parameter diffusion model developed by Freepik and Fal.
It has been built through knowledge distillation from F Lite and, as such, it delivers image quality comparable to its larger sibling while being faster and more memory-efficient. It also preserves the same copyright-safe, SFW generation characteristics as F Lite.
Usage
Try F Lite 7B instantly through our interactive demo on Hugging Face.
F Lite 7B works with both the diffusers library and ComfyUI. For details, see the F Lite GitHub repository.
Technical Report
Read the technical report to learn more about the model architecture, training recipe and evaluation.
Limitations and Bias
- Like its teacher, the model can generate malformations.
- Text rendering capabilities remain limited.
- Despite careful curation, societal and aesthetic biases can still emerge.
Recommendations
- Longer, descriptive prompts improve quality; very short prompts may yield sub-optimal images.
- Generate images above the megapixel mark; low resolutions reduce fidelity.
Acknowledgements
This model relies on high-quality components including T5 XXL for prompt encoding and the Flux Schnell VAE for efficient latent-space compression.
License
The F Lite 7B weights are released under the permissive CreativeML Open RAIL-M license. Dependencies inherit their respective licenses (Apache 2.0 for T5 XXL and Flux Schnell VAE).
Citation
@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}
}
- Downloads last month
- 17