Flux.1 Dedistilled Mix Tuned fp8

Introduction

The FLUX.1-Dedistilled-Mix-Tuned-FP8 model is a high-performance text-to-image model designed for fast and efficient image generation. It is specifically optimized to achieve a balance between image quality, detail, realism, and style diversity. This model is a derivative of the FLUX.1 Schnell and Dev models, incorporating various fine-tuning and merging techniques to enhance its capabilities.

Architecture

The model builds upon the FLUX.1 Schnell architecture, integrating components from LibreFLUX and other Flux models. It utilizes Block Patcher and ComfyUI tools to refine its performance. The model also supports GGUF quantization versions such as Q8_0, Q5_1, and Q4_1, though it is recommended to use the FP8 version for optimal results.

Training

The FLUX.1-Dedistilled-Mix-Tuned-FP8 has been fine-tuned using ComfyUI and other specialized tools to enhance its image generation capabilities. It is recommended to use 4-8 steps for the FLUX.1 Schnell-based model and 6-10 steps for the FLUX.1 Dev-based model to achieve the best results.

Guide: Running Locally

1. Install Required Software: Ensure you have the necessary libraries, such as the Hugging Face transformers and diffusers.
2. Download the Model: Obtain the FLUX.1-Dedistilled-Mix-Tuned-FP8 model files from Hugging Face.
3. Set Up Environment: Utilize the recommended CLIP and Text Encoder models for better prompt guidance.
4. Run the Model: Execute the model using a supported platform. For GGUF versions, employ the UNET Loader(GGUF) node from city96's repository.

For optimal performance, consider using cloud GPUs available from providers like AWS, Google Cloud, or Azure.

License

The model weights fall under the FLUX.1 [dev] Non-Commercial License. Please refer to the license document for specific usage rights and restrictions.