---
language: en
license:  apache-2.0
tags:
  - text-to-image
  - diffusion
  - mflux
  - development
datasets:
  - custom
---

# FLUX.1-dev-mflux-4bit

[![Hugging Face](https://img.shields.io/badge/🤗%20Hugging%20Face-FLUX.1--dev--mflux--4bit-blue)](https://huggingface.co/dhairyashil/FLUX.1-dev-mflux-4bit)

![comparison_output](comparison.png)

A quantized version of the [FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev) text-to-image model, implemented using the [mflux](https://github.com/filipstrand/mflux) (version 0.6.2) quantization approach. 

## Overview

This repository contains the 4-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 4-bit Quantization

- **Reduced Memory Usage**: ~75% 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 (4-bit quantized)
- `text_encoder_2/`: Secondary text encoder (4-bit quantized)
- `tokenizer/`: CLIP tokenizer configuration and vocabulary
- `tokenizer_2/`: Secondary tokenizer configuration
- `transformer/`: Main diffusion model components (4-bit quantized)
- `vae/`: Variational autoencoder for image encoding/decoding (4-bit quantized)

## Usage

### Requirements

- Python 
- PyTorch 
- Transformers 
- Diffusers 
- [mflux](https://github.com/filipstrand/mflux) library (for 8-bit model support)

### Installation

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

### Example Usage

```bash
# export path for mflux
% mflux-generate \        
mflux-generate \
    --path "dhairyashil/FLUX.1-dev-mflux-4bit" \
    --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 but 4-bit version may show some quality loss.

A [8-bit model](https://huggingface.co/dhairyashil/FLUX.1-dev-mflux-8bit) 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](https://huggingface.co/black-forest-labs) for creating the original FLUX.1 model family
- [Filip Strand](https://github.com/filipstrand) 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](https://huggingface.co/black-forest-labs/FLUX.1) for licensing information.