README.md

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

# Inpainting with Perturbed-Attention Guidance

[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 StableDiffusionPipeline to support inpainting with Perturbed-Attention Guidance (PAG). Please refer to "Inpainting" section of an [official document](https://huggingface.co/docs/diffusers/using-diffusers/inpaint) for details.

## Quickstart

Loading Custom Piepline:

```
from diffusers import StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-inpainting",
    custom_pipeline="hyoungwoncho/sd_perturbed_attention_guidance_inpaint",
    torch_dtype=torch.float16,
    safety_checker=None
)

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

Inpainting with PAG:

```
output = pipe(
    prompts,
    image=init_image,
    mask_image=mask_image,
    num_inference_steps=50,
    guidance_scale=0.0,
    pag_scale=3.0,
    pag_applied_layers_index=['u0']
).images[0]
```

## Parameters

guidance_scale : gudiance scale of CFG (ex: 7.5)

pag_scale : gudiance scale of PAG (ex: 3.0)

pag_applied_layers_index : index of the layer to apply perturbation (ex: ['u0'])