README_original.md

---
license: other
license_name: flux-1-dev-non-commercial-license
license_link: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md
base_model:
- ashen0209/Flux-Dev2Pro
- black-forest-labs/FLUX.1-schnell
tags:
- text-to-image
- flux
- merge
widget:
- text: Example 1
  output:
    url: images/0015.webp
- text: Example 2
  output:
    url: images/0022.webp
- text: Example 3
  output:
    url: images/0012.webp
- text: Example 4
  output:
    url: images/0014.webp
- text: Example 5
  output:
    url: images/0033.webp
- text: Example 6
  output:
    url: images/0032.webp
---

# Flux D+S F8 Diffusers

A + B merge, meant for gradio inference, possibly build adapters on top with [diffusers](https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/train_dreambooth_lora_flux.py).

<Gallery />

## Inference

[Instruction](https://huggingface.co/twodgirl/flux-dev-fp8-e4m3fn-diffusers)

This model needs 10 steps, otherwise the images remain blurry.

## ComfyUI

I convert all my checkpoints locally. There are 5+ file formats out there, they take up too much space.

Download the contents of this repo, then enter the directory. A script will convert the diffusers format to something that is compatible with ComfyUI.

```
cd flux-devpro-schnell-merge-fp8-e4m3fn-diffusers
pip install safetensors torch
python flux_devpro_ckpt.py transformer/diffusion_pytorch_model.safetensors flux-devpro-schnell.safetensors
mv flux-devpro-schnell.safetensors path/to/your/ComfyUI/models/unet/flux-devpro-schnell.safetensors
```

In the workflow, the LoadDiffusionModel node should take two values: flux-devpro-schnell.safetensors, fp8_e4m3fn.

Use of this code requires citation and attribution to the author via a link to their Hugging Face profile in all resulting work.

## Disclaimer

Sharing, reusing the model weights requires a link back to authors and their Hugging Face profile. The source models were uploaded by Ashen0209 and BFL.