README.md

---
license: apache-2.0
language:
- en
library_name: diffusers
pipeline_tag: image-to-image
---

# InstantIR Model Card

<!-- > **InstantIR: Blind Image Restoration with Instant Generative Reference**<br>
> Jen-Yuan Huang<sup>1,2</sup>, Haofan Wang<sup>2</sup>, Qixun Wang<sup>2</sup>, Xu Bai<sup>2</sup>, Hao Ai<sup>2</sup>, Peng Xing<sup>2</sup>, Jen-Tse Huang<sup>3</sup> <br>
> <sup>1</sup>Peking University, <sup>2</sup>InstantX Team, <sup>3</sup>The Chinese University of Hong Kong -->

<a href='https://arxiv.org/abs/2410.06551'><img src='https://img.shields.io/badge/arXiv-b31b1b.svg'></a>
<a href='https://jy-joy.github.io/InstantIR'><img src='https://img.shields.io/badge/Website-informational'></a>
<a href='https://github.com/JY-Joy/InstantIR'><img src='https://img.shields.io/badge/Github-gray'></a>

> **InstantIR** is a novel single-image restoration model designed to resurrect your damaged images, delivering extrem-quality yet realistic details. You can further boost **InstantIR** performance with additional text prompts, even achieve customized editing!

<div  align="center">
<img src='assets/teaser_figure.png'>
</div>


## Usage

### 1. Clone the github repo
```sh
git clone https://github.com/JY-Joy/InstantIR.git
cd InstantIR
```

### 2. Download model weights
You can directly download InstantIR weights in this repository, or
you can download them using python script:

```python
from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="InstantX/InstantIR", filename="models/adapter.pt", local_dir="./models")
hf_hub_download(repo_id="InstantX/InstantIR", filename="models/aggregator.pt", local_dir="./models")
hf_hub_download(repo_id="InstantX/InstantIR", filename="models/previewer_lora_weights.bin", local_dir="./models")
```

### 3. Load InstantIR with 🧨 diffusers

```python
# !pip install opencv-python transformers accelerate
import torch
from PIL import Image

import diffusers
from diffusers import DDPMScheduler, StableDiffusionXLPipeline
from diffusers.utils import load_image
from schedulers.lcm_single_step_scheduler import LCMSingleStepScheduler

from transformers import AutoImageProcessor, AutoModel

from module.ip_adapter.utils import load_ip_adapter_to_pipe, revise_state_dict, init_ip_adapter_in_unet
from module.ip_adapter.resampler import Resampler
from module.aggregator import Aggregator
from pipelines.sdxl_instantir import InstantIRPipeline

# prepare 'dinov2'
image_encoder = AutoModel.from_pretrained('facebook/dinov2-large')
image_processor = AutoImageProcessor.from_pretrained('facebook/dinov2-large')

# prepare models under ./checkpoints
dcp_adapter = f'./models/adapter.pt'
previewer_lora_path = f'./models'
instantir_path = f'./models/aggregator.pt'

# load SDXL
sdxl = StableDiffusionXLPipeline.from_pretrained('stabilityai/stable-diffusion-xl-base-1.0', torch_dtype=torch.float16)

# load adapter
image_proj_model = Resampler(
    embedding_dim=image_encoder.config.hidden_size,
    output_dim=sdxl.unet.config.cross_attention_dim,
)
init_ip_adapter_in_unet(
    sdxl.unet,
    image_proj_model,
    dcp_adapter,
)

pipe = InstantIRPipeline(
    sdxl.vae, sdxl.text_encoder, sdxl.text_encoder_2, sdxl.tokenizer, sdxl.tokenizer_2,
    sdxl.unet, sdxl.scheduler, feature_extractor=image_processor, image_encoder=image_encoder,
)
pipe.cuda()

# load previewer lora
pipe.prepare_previewers(previewer_lora_path)
pipe.unet.to(dtype=torch.float16)
pipe.scheduler = DDPMScheduler.from_pretrained('stabilityai/stable-diffusion-xl-base-1.0', subfolder="scheduler")
lcm_scheduler = LCMSingleStepScheduler.from_config(pipe.scheduler.config)

# load aggregator weights
pretrained_state_dict = torch.load(instantir_path)
pipe.aggregator.load_state_dict(pretrained_state_dict)
pipe.aggregator.to(dtype=torch.float16)
```

Then, you can restore your broken images with:

```python
# load a broken image
image = Image.open('path/to/your-image').convert("RGB")

# InstantIR restoration
image = pipe(
    prompt='',
    image=image,
    ip_adapter_image=[image],
    negative_prompt='',
    guidance_scale=7.0,
    previewer_scheduler=lcm_scheduler,
    return_dict=False,
)[0]
```

For more details including text-guided enhancement/editing, please refer to our [GitHub repository](https://github.com/JY-Joy/InstantIR). 

<!-- ## Usage Tips
1. If you're not satisfied with the similarity, try to increase the weight of "IdentityNet Strength" and "Adapter Strength".
2. If you feel that the saturation is too high, first decrease the Adapter strength. If it is still too high, then decrease the IdentityNet strength.
3. If you find that text control is not as expected, decrease Adapter strength.
4. If you find that realistic style is not good enough, go for our Github repo and use a more realistic base model. -->

## Examples

<div  align="center">
<img src='assets/qualitative_real.png'>
</div>

<div  align="center">
<img src='assets/outdomain_preview.png'>
</div>

## Disclaimer

This project is released under Apache License and aims to positively impact the field of AI-driven image generation. Users are granted the freedom to create images using this tool, but they are obligated to comply with local laws and utilize it responsibly. The developers will not assume any responsibility for potential misuse by users.

## Citation
```bibtex
@article{huang2024instantir,
  title={InstantIR: Blind Image Restoration with Instant Generative Reference},
  author={Huang, Jen-Yuan and Wang, Haofan and Wang, Qixun and Bai, Xu and Ai, Hao and Xing, Peng and Huang, Jen-Tse},
  journal={arXiv preprint arXiv:2410.06551},
  year={2024}
}
```