Diffusion Inversion Methods

This repository is a collection of diffusion inversion methods. We implemented all mentioned methods in the repository using diffusers.

NOTE that this repository is only build for inversion, applications to downstream tasks (e.g. image editing, image variation, image inpainting, and rare concept generation, etc) are not included.

1. Installation

Create and activate a conda environment:

conda create -n diffusion-inversion python=3.10

conda activate diffusion-inversion

Clone this Repository:

git clone https://github.com/wmchen/diffusion-inversion.git
cd diffusion-inversion

Install PyTorch:

pip install -r requirements/torch.txt

Install other packages:

pip install -r requirements/build.txt

2. Getting Started

Please refer to examples/ for a quick start.

3. Supported Methods

3.1 UNet-based Diffusion Models

Method	Publication	Paper	Official repo.	Ours imp.
DDIM Inversion	ICLR 2021	paper		ours
Negative Prompt Inversion (NPI)	ArXiv 2023	paper		ours
Null-Text Inversion (NTI)	CVPR 2023	paper	official	ours
Exact Diffusion Inversion via Coupled Transformations (EDICT)	CVPR 2023	paper	official	ours
Accelerated Iterative Diffusion Inversion (AIDI)	ICCV 2023 Oral	paper		ours
Prompt Tuning Inversion (PTI)	ICCV 2023	paper		ours
Real-world Image Variation by ALignment (RIVAL)	NeurIPS 2023 Spotlight	paper	official	ours
Fixed-Point Inversion (FPI)	ArXiv 2023	paper	official	ours
On Exact Inversion of DPM-solvers	CVPR 2024	paper	official	ours
Tuning-free Inversion-enhanced Control (TIC)	AAAI 2024	paper		ours
Bi-Directional Integration Approximation (BDIA)	ECCV 2024 Oral	paper	official	ours
ReNoise	ECCV 2024	paper	official	ours
Bidirectional Explicit Linear Multi-step (BELM)	NeurlPS 2024	paper	official	ours
Guided Newton Raphson Inversion (GNRI)	ICLR 2025	paper	official	ours

All methods above support Stable Diffusion v1, v2 and SDXL in our implementations. The following versions of diffusion model are fully tested:

• stable-diffusion-v1-5
• stable-diffusion-v2-1
• stable-diffusion-xl-base-1.0

3.2 DiT-based Diffusion Models

Method	Publication	Paper	Official repo.	Ours imp.
RF-Inversion	ICLR 2025	paper	official	ours
RF-Solver	ICML 2025	paper	official	ours
FireFlow	ICML 2025	paper	official	ours

CLAIM: The implementation of RF-Inversion is copy from diffusers.

4. Citation

If you use this codebase in your research, please cite our repository and the corresponding papers.

@misc{chen2024inversion,
  title = {Diffusion Inversion Methods},
  author = {Chen, Weiming},
  howpublished = {https://github.com/wmchen/diffusion-inversion},
  year = {2024},
}

5. Contributors

We appreciate all the contributors who add new features or fix bugs, as well as the users who offer valuable feedback. We welcome all contributors, feel free to create an issue or file a pull request and join us! ❤️

Weiming Chen 🤔 💻 📆 🔬	liuqifan67 🤔 🔬	llsysysy 🤔 🔬	TANG 🤔 🔬	nkdailab 💵
Add your contributions

6. License

This project is intended for research use only, licensed under the Apache-2.0 license.