SemBind

KO | EN

gitlite — search

#chatbot #devops #pytorch #git #codex #react-native #ai-agents #typescript #llm #cross-platform

(★ 9)

Official implementation of "SemBind: Binding Diffusion Watermarks to Semantics Against Black-Box Forgery Attacks" (ICML 2026).

File Explorer

ZIP Download

- inverse_stable_diffusion.py
- modified_stable_diffusion.py
- optim_utils.py
- prc.py
- pseudogaussians.py
- train_new.py
.gitignore
environment.yml
image_utils.py
inversion.py
LICENSE
model.py
README.md
semantimark.ipynb
semantimark_demo.png
utils.py

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

Official implementation of **SemBind: Binding Diffusion Watermarks to Semantics Against Black-Box Forgery Attacks**, accepted to **ICML 2026**.

SemBind is a semantic binding framework for diffusion watermarks. It binds latent watermark signals to image semantics through a learned semantic masker, improving resistance against black-box forgery attacks while maintaining compatibility with existing latent-based watermarking schemes.

## Overview

Black-box forgery attacks can transfer or forge invisible diffusion watermarks by exploiting watermarked outputs from a target model. SemBind mitigates this threat by introducing a semantic masker that extracts a semantic-dependent binary code from an auxiliary clean image. This code is then used to modulate the watermark embedding process, binding the watermark to the semantic content of the generated image.

The semantic masker is trained to produce:

* consistent codes for images generated from the same or semantically similar prompts;
* separable codes for images from different prompts;
* robust codes under common image distortions.

## Repository Status

Code and instructions are being released progressively.

Currently, this repository includes a demo notebook for visualizing how the semantic masker works.

## Semantic Masker Demo

Please refer to:

```text
semantimark.ipynb
```

This notebook demonstrates how the semantic masker extracts semantic-dependent binary hash codes from generated images.

In the demo, we generate multiple images from the same prompt using Stable Diffusion 2.1 Base, apply several image distortions, and then compute semantic hash codes using a DINOv2 feature extractor and the trained semantic masker.

The main workflow is:

1. Load Stable Diffusion 2.1 Base.
2. Load the DINOv2 image encoder.
3. Load the trained semantic masker.
4. Generate multiple images from one prompt.
5. Apply image distortions to selected images.
6. Extract DINOv2 CLS features.
7. Compute binary semantic codes using the semantic masker.
8. Visualize images together with their corresponding hash codes.

A simplified example is shown below:

```python
# Load DINOv2 model
dino_processor = AutoImageProcessor.from_pretrained(dinov2_path)
dino_model = AutoModel.from_pretrained(dinov2_path).to(device)
dino_model.eval()

# Load semantic masker
hashdim = 1024
featdim = 8192
hash_model_path = './model/@3M_attn_hash_model_180_160_hashdim_1024.pth'

hash_model = AttnSupConHashingModel(
    input_dim=1536,
    feat_dim=featdim,
    hash_dim=hashdim
).to(device)

hash_model.load_state_dict(torch.load(hash_model_path, map_location=device))
hash_model.eval()

# Compute semantic hash code
inp = dino_processor(images=img, return_tensors='pt').to(device)

with torch.no_grad():
    cls = dino_model(**inp).last_hidden_state[:, 0, :]
    cls = F.normalize(cls, dim=1)
    bits = hash_model(cls, stage='binarize', s=12.0)
```

The notebook visualizes original and distorted images together with their extracted semantic hash codes. This allows users to observe how the semantic masker produces stable semantic codes for images sharing the same semantic content.

![Semantic masker demo](https://raw.githubusercontent.com/XinZhang1999/SemBind/main/semantimark_demo.png)

## Environment

We recommend creating a Conda environment before running the code.

```bash
conda env create -f environment.yml
conda activate sembind
```

If you use a different environment name, please replace `sembind` accordingly.

## Model Weights

The pretrained semantic masker checkpoint can be downloaded from Google Drive:

```text
@3M_attn_hash_model_180_160_hashdim_1024.pth
```

Download link:

```text
https://drive.google.com/file/d/1-CX1LerxBNbJDxlBrneQCcwBBNnn3iCQ/view?usp=sharing
```

Please place the downloaded checkpoint under:

```text
./model/
```

The expected file structure is:

```text
SemBind/
├── model/
│   └── @3M_attn_hash_model_180_160_hashdim_1024.pth
├── semantimark.ipynb
└── ...
```

The checkpoint path used in `semantimark.ipynb` is:

```python
hash_model_path = './model/@3M_attn_hash_model_180_160_hashdim_1024.pth'
```

## Data and Model Paths

The demo notebook may contain local paths such as:

```text
/home/goodbacteria/models/stable-diffusion-2-1-base
/home/goodbacteria/models/dinov2-giant
/home/goodbacteria/models/coco/captions_val2017.json
```

Please modify these paths according to your local environment.

For example, you may set:

```python
model_id = "/path/to/stable-diffusion-2-1-base"
dinov2_path = "/path/to/dinov2-giant"
hash_model_path = "./model/@3M_attn_hash_model_180_160_hashdim_1024.pth"
```

## Citation

If you find this repository useful, please consider citing our paper:

```bibtex
@inproceedings{zhang2026sembind,
  title     = {SemBind: Binding Diffusion Watermarks to Semantics Against Black-Box Forgery Attacks},
  author    = {Zhang, Xin and others},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year      = {2026}
}
```

The official BibTeX will be updated after the ICML 2026 proceedings are released.

## License

This project is released under the MIT License.

SemBind

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

Download README