FLUX.1-dev-mflux-8bit

A quantized version of the FLUX.1-dev text-to-image model, implemented using the mflux (version 0.6.2) quantization approach.

Overview

This repository contains the 8-bit quantized FLUX.1 model, which significantly reduces the memory footprint while maintaining most of the generation quality. The quantization was performed using the mflux.

Benefits of 8-bit Quantization

Reduced Memory Usage: ~50% reduction in memory requirements compared to the original model
Faster Loading Times: Smaller model size means quicker initialization
Lower Storage Requirements: Significantly smaller disk footprint
Accessibility: Can run on consumer hardware with limited VRAM
Minimal Quality Loss: Maintains nearly identical output quality to the original model

Model Structure

This repository contains the following components:

text_encoder/: CLIP text encoder (8-bit quantized)
text_encoder_2/: Secondary text encoder (8-bit quantized)
tokenizer/: CLIP tokenizer configuration and vocabulary
tokenizer_2/: Secondary tokenizer configuration
transformer/: Main diffusion model components (8-bit quantized)
vae/: Variational autoencoder for image encoding/decoding (8-bit quantized)

Usage

Requirements

Python
PyTorch
Transformers
Diffusers
mflux library (for 8-bit model support)

Installation

pip install torch diffusers transformers accelerate
uv tool install mflux # check mflux README for more details

Example Usage

# export path for mflux
% mflux-generate \
    --path "dhairyashil/FLUX.1-dev-mflux-8bit" \
    --model dev \
    --steps 50 \
    --seed 2 \
    --height 1920 \
    --width 1024 \
    --prompt "hot chocolate dish on decorated table"

Comparison Output

The images generated from above prompt for different models are shown at the top.

fp16 and 8-bit results look visibly almost the same, with the 8-bit version maintaining excellent quality while using significantly less memory.

A 4-bit development model may also be available for testing, though with more noticeable quality difference.

Performance Comparison

Model Version	Memory Usage	Inference Speed	Quality
Original FP16	~36 GB	Base	Base
8-bit Quantized	~18 GB	Nearly identical	Nearly identical
4-bit Quantized	~9 GB	Nearly identical	Moderately reduced

Other Highlights

Very minimal quality degradation compared to the original model
Nearly identical inference speed
Rare artifacts that are generally imperceptible in most use cases

Acknowledgements

Black Forest Labs for creating the original FLUX.1 model family
Filip Strand for developing the mflux quantization methodology
The Hugging Face team for their Diffusers and Transformers libraries
All contributors to the development version for their testing and improvements

License

This model inherits the license of the original FLUX.1 model. Please refer to the original model repository for licensing information.