README.md

---
base_model:
- SherryXTChen/LatentDiffusionDINOv2
datasets:
- timbrooks/instructpix2pix-clip-filtered
- SherryXTChen/InstructCLIP-InstructPix2Pix-Data
language:
- en
license: apache-2.0
pipeline_tag: image-to-image
library_name: diffusers
tags:
- model_hub_mixin
- pytorch_model_hub_mixin
---

# InstructCLIP: Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning (CVPR 2025)

This model has been pushed to the Hub using the [PytorchModelHubMixin](https://huggingface.co/docs/huggingface_hub/package_reference/mixins#huggingface_hub.PyTorchModelHubMixin) integration.
The model is based on the paper [Instruct-CLIP: Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning](https://huggingface.co/papers/2503.18406).

[Arxiv](http://arxiv.org/abs/2503.18406) | [Image Editing Model](https://huggingface.co/SherryXTChen/InstructCLIP-InstructPix2Pix) | [Data Refinement Model](https://huggingface.co/SherryXTChen/Instruct-CLIP) | [Data](https://huggingface.co/datasets/SherryXTChen/InstructCLIP-InstructPix2Pix-Data)


## Capabilities

<p align="center">
  <img src="https://raw.githubusercontent.com/SherryXTChen/Instruct-CLIP/refs/heads/main/assets/teaser_2.png" alt="Figure 2" width="50%">
</p>

## Installation
```
pip install -r requirements.txt
```

## Edit Instruction Refinement Inference

```python
from PIL import Image
import torch
from torchvision import transforms

from model import InstructCLIP
from utils import get_sd_components, normalize

parser = argparse.ArgumentParser(description="Simple example of estimating edit instruction from image pair")
parser.add_argument(
    "--pretrained_instructclip_name_or_path",
    type=str,
    default="SherryXTChen/Instruct-CLIP",
    help=(
        "instructclip pretrained checkpoints"
    ),
)
parser.add_argument(
    "--pretrained_model_name_or_path",
    type=str,
    default="runwayml/stable-diffusion-v1-5",
    help=(
        "sd pretrained checkpoints"
    ),
)
parser.add_argument(
    "--input_path",
    type=str,
    default="assets/1_input.jpg",
    help=(
        "Input image path"
    )
)
parser.add_argument(
    "--output_path",
    type=str,
    default="assets/1_output.jpg",
    help=(
        "Output image path"
    )
)
args = parser.parse_args()
device = "cuda"
    
# load model for edit instruction estimation
model = InstructCLIP.from_pretrained("SherryXTChen/Instruct-CLIP")
model = model.to(device).eval()

# load model to preprocess/encode image to latent space
tokenizer, _, vae, _, _ = get_sd_components(args, device, torch.float32)

# prepare image input
transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.5], std=[0.5]),
])
image_list = [args.input_path, args.output_path]
image_list = [
    transform(Image.open(f).resize((512, 512))).unsqueeze(0).to(device) 
    for f in image_list
]

with torch.no_grad():
    image_list = [vae.encode(x).latent_dist.sample() * vae.config.scaling_factor for x in image_list]
    
    # get image feature
    zero_timesteps = torch.zeros_like(torch.tensor([0])).to(device) 
    img_feat = model.get_image_features(
        inp=image_list[0], out=image_list[1], inp_t=zero_timesteps, out_t=zero_timesteps)
    img_feat = normalize(img_feat)
    
    # get edit instruction
    pred_instruct_input_ids = model.text_decoder.infer(img_feat[:1])[0]
    pred_instruct = tokenizer.decode(pred_instruct_input_ids, skip_special_tokens=True)
    print(pred_instruct)  # as a 3 d sculpture
```

## Citation
```bibtex
@misc{chen2025instructclipimprovinginstructionguidedimage,
      title={Instruct-CLIP: Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning}, 
      author={Sherry X. Chen and Misha Sra and Pradeep Sen},
      year={2025},
      eprint={2503.18406},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.18406}, 
}
```