---
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
language:
- en
base_model: black-forest-labs/FLUX.1-dev
library_name: diffusers
tags:
- Text-to-Image
- FLUX
- Stable Diffusion
pipeline_tag: text-to-image
---

<div style="display: flex; justify-content: center; align-items: center;">
  <img src="./images/images_alibaba.png" alt="alibaba" style="width: 20%; height: auto; margin-right: 5%;">
  <img src="./images/images_alimama.png" alt="alimama" style="width: 20%; height: auto;">
</div>

本仓库包含了由阿里妈妈创意团队开发的基于[FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev)模型的8步蒸馏版。

# 介绍

该模型是基于FLUX.1-dev模型的8步蒸馏版lora。我们使用特殊设计的判别器来提高蒸馏质量。该模型可以用于T2I、Inpainting controlnet和其他FLUX相关模型。建议guidance_scale=3.5和lora_scale=1。我们的更低步数的版本将在后续发布。

- Text-to-Image.

![](./images/T2I.png)

- 配合[alimama-creative/FLUX.1-dev-Controlnet-Inpainting-Beta](https://huggingface.co/alimama-creative/FLUX.1-dev-Controlnet-Inpainting-Beta)。我们模型可以很好地适配Inpainting controlnet，并与原始输出保持相似的结果。

![](./images/inpaint.png)

# 使用指南
## diffusers
该模型可以直接与diffusers一起使用

```python
import torch
from diffusers.pipelines import FluxPipeline

model_id = "black-forest-labs/FLUX.1-dev"
adapter_id = "alimama-creative/FLUX.1-Turbo-Alpha"

pipe = FluxPipeline.from_pretrained(
  model_id,
  torch_dtype=torch.bfloat16
)
pipe.to("cuda")

pipe.load_lora_weights(adapter_id)
pipe.fuse_lora()

prompt = "A DSLR photo of a shiny VW van that has a cityscape painted on it. A smiling sloth stands on grass in front of the van and is wearing a leather jacket, a cowboy hat, a kilt and a bowtie. The sloth is holding a quarterstaff and a big book."
image = pipe(
            prompt=prompt,
            guidance_scale=3.5,
            height=1024,
            width=1024,
            num_inference_steps=8,
            max_sequence_length=512).images[0]
```

## comfyui

- 文生图加速链路: [点击这里](./workflows/t2I_flux_turbo.json)
- Inpainting controlnet 加速链路: [点击这里](./workflows/alimama_flux_inpainting_turbo_8step.json)


# 训练细节

该模型在1M公开数据集和内部源图片上进行训练，这些数据美学评分6.3+而且分辨率大于800。我们使用对抗训练来提高质量，我们的方法将原始FLUX.1-dev transformer固定为判别器的特征提取器，并在每个transformer层中添加判别头网络。在训练期间，我们将guidance scale固定为3.5，并使用时间偏移量3。

混合精度: bf16

学习率: 2e-5

批大小: 64

训练分辨率: 1024x1024