README.md

---
base_model:
- CompVis/stable-diffusion-v1-4
pipeline_tag: text-to-image
tags:
- art
---
<h1 align="center">The Superposition of Diffusion Models Using the Itô Density Estimator: <em>Pipeline</em></h1>

 <p align="center">
 <a href="https://arxiv.org/abs/2412.17762"><img src="https://img.shields.io/badge/Arxiv-2412.17762-red?style=for-the-badge&logo=Arxiv" alt="arXiv"/></a>
</p>

This pipeline shows how to superimpose different text prompts from [Stable Diffusion v1-4](https://huggingface.co/CompVis/stable-diffusion-v1-4) based the paper [The Superposition of Diffusion Models Using the Itô Density Estimator](https://www.arxiv.org/abs/2412.17762).

<p align="center">
<img src="https://huggingface.co/superdiff/superdiff-sd-v1-4/resolve/main/superdiff_small.gif" alt="drawing" style="width:500px;">
</p>

## Requirements

This pipeline can be run with the following packages & versions:
- `PyTorch 2.5.1`
- `Diffusers 0.32.1`
- `Accelerate 1.2.1`
- `Transformers 4.47.1`

You can install these with:

```
pip install torch
pip install diffusers accelerate transformers
```


## Example usage

```
from PIL import Image
from diffusers import DiffusionPipeline

pipeline = DiffusionPipeline.from_pretrained("superdiff/superdiff-sd-v1-4", custom_pipeline='pipeline', trust_remote_code=True)
output = pipeline("a flamingo", "a candy cane", seed=1, num_inference_steps=1000, batch_size=1)

image = Image.fromarray(output[0].cpu().numpy())
image.save("superdiff_output.png")
```

Arguments that can be set by user in `pipeline()`:

- `prompt_1` [required]: text prompt describing first concept to superimpose (e.g. "a flamingo")
- `prompt_2`[required]: text prompt describing second concept to superimpose (e.g. "a candy cane")
- `seed`[optional: default=None]: seed for random noise generator for reproducibility; for non-deterministic outputs, set to `None`
- `num_inference_steps`[optional: default=1000]: number of denoising steps (we recommend 1000!)
- `batch_size` [optional: default=1]: batch size
- `lift` [optional: default=0.0]: bias value that favours generation towards one prompt over the other
- `guidance_scale` [optional: default=7.5]: scale for classifier-free guidance
- `height`, `width` [optional: default=512]: height and width of generated images

To replicate images from Section 4.2 of the paper, you can use the following:

```
image = pipeline(prompt_1, prompt_2, seed=1, num_inference_steps=1000, batch_size=20, lift=0.0, guidance_scale=7.5)
```

(Note: the runtime for a batch size of 1 on an NVIDIA A40 GPU is around 3 mins 30 sec.)

## Citation

**BibTeX:**

```
@article{skreta2025superposition,
  title={The Superposition of Diffusion Models Using the It$\backslash$\^{} o Density Estimator},
  author={Skreta, Marta and Atanackovic, Lazar and Bose, Avishek Joey and Tong, Alexander and Neklyudov, Kirill},
  journal={International Conference on Learning Representations},
  year={2025}
}
```