LibreFLUX
| Property | Value |
|---|---|
| License | Apache 2.0 |
| Training Compute | ~1,500 H100 hours equivalent |
| Library | Diffusers |
| Pipeline Type | Text-to-Image |
What is LibreFLUX?
LibreFLUX is an Apache 2.0 licensed de-distilled version of FLUX.1-schnell, designed to provide full functionality with restored classifier-free guidance and complete T5 context length support. The model was trained using approximately 1,500 H100-equivalent hours and implements attention masking for improved token utilization.
Implementation Details
The model employs several innovative technical approaches, including LoKr parameter-efficient fine-tuning with 3.2B parameters, beta timestep scheduling, and multi-rank stratified sampling. It was trained on about 0.5 million high-resolution images with diverse captions.
- Full 512 token context length (upgraded from original 256)
- Attention masking implementation for better token utilization
- Restored classifier-free guidance functionality
- Quantization support for lower VRAM usage
Core Capabilities
- High-quality text-to-image generation
- Support for both short and long text prompts
- Efficient inference with int8 quantization option
- Easy integration with Fine-tuning frameworks
Frequently Asked Questions
Q: What makes this model unique?
LibreFLUX stands out by offering a fully open-source, Apache 2.0 licensed version of FLUX with restored functionality and improved token handling, making it suitable for both commercial use and further fine-tuning.
Q: What are the recommended use cases?
The model excels at general text-to-image generation tasks and works best with CFG scale of 2.0 to 5.0. It's particularly suitable for commercial applications requiring an open-source license and scenarios needing long text prompt processing.
