README.md

---
language:
- en
pipeline_tag: unconditional-image-generation
tags:
- Diffusion Models
- Stable Diffusion
- Perturbed-Attention Guidance
- PAG
---

# Perturbed-Attention Guidance for SDXL
![image/jpeg](uncond_generation_pag.jpg)
![image/jpeg](cfgpag.jpg)

[Project](https://ku-cvlab.github.io/Perturbed-Attention-Guidance/) / [arXiv](https://arxiv.org/abs/2403.17377) / [GitHub](https://github.com/KU-CVLAB/Perturbed-Attention-Guidance)

This repository is based on [Diffusers](https://huggingface.co/docs/diffusers/index). The pipeline is a modification of StableDiffusionXLPipeline to add Perturbed-Attention Guidance (PAG).

The original Perturbed-Attention Guidance for unconditional models and SD1.5 by [Hyoungwon Cho](https://huggingface.co/hyoungwoncho) is availiable at [hyoungwoncho/sd_perturbed_attention_guidance](https://huggingface.co/hyoungwoncho/sd_perturbed_attention_guidance)

## Quickstart

Loading Custom Pipeline:

```py
from diffusers import StableDiffusionXLPipeline

pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    custom_pipeline="multimodalart/sdxl_perturbed_attention_guidance",
    torch_dtype=torch.float16
)

device="cuda"
pipe = pipe.to(device)
```

Unconditional sampling with PAG:

```py
output = pipe(
        "",
        num_inference_steps=50,
        guidance_scale=0.0,
        pag_scale=5.0,
        pag_applied_layers=['mid']
    ).images
```

Sampling with PAG and CFG:

```py
output = pipe(
        "the spirit of a tamagotchi wandering in the city of Vienna",
        num_inference_steps=25,
        guidance_scale=4.0,
        pag_scale=3.0,
        pag_applied_layers=['mid']
    ).images
```

## Parameters

`guidance_scale` : gudiance scale of CFG (ex: `7.5`)

`pag_scale` : gudiance scale of PAG (ex: `4.0`)

`pag_applied_layers`: layer to apply perturbation (ex: ['mid'])

`pag_applied_layers_index` : index of the layers to apply perturbation (ex: ['m0', 'm1'])

## Stable Diffusion XL Demo

Soon