PreciseInv

This repository contains the official implementation of our NeurIPS 2025 Poster paper:

Precise Diffusion Inversion: Towards Novel Samples and Few-Step Models
Jing Zuo · Luoping Cui · Chuang Zhu · Yonggang Qi

---

Installation

conda create -n preciseinv python==3.11.*
conda activate preciseinv
conda install pytorch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 pytorch-cuda=12.4 -c pytorch -c nvidia
pip install diffusers transformers lpips scikit-image matplotlib

---

Quick Start

python inversion.py \
    --input_images input_images/dog.jpg \
    --model_name SD14 \
    --scheduler_name DDPM \
    --num_inference_steps 2 \
    --batch_size 1 \
    --eta 1e-2 \
    --learning_rate 0.1 \
    --save_dir outputs

---

Supported Models

Model Name	Scheduler	Status
SD14	DDIM / DDPM	✅ Supported
SD15	DDIM / DDPM	✅ Supported
SD21	DDIM / DDPM	✅ Supported
SDXL	DDIM / DDPM	✅ Supported
LCM-SD15	DDIM / DDPM	✅ Supported
LCM-SDXL	DDIM / DDPM	✅ Supported
SDXL-Turbo	DDIM / DDPM	✅ Supported
SD3	Euler	✅ Supported
FLUX	Euler	⚠️ Planned
ShortCut	Euler	⚠️ Planned
MeanFlow	Euler	⚠️ Planned

---

Citation

@inproceedings{zuo2025precise,
  title     = {Precise Diffusion Inversion: Towards Novel Samples and Few-Step Models},
  author    = {Zuo, Jing and Cui, Luoping and Zhu, Chuang and Qi, Yonggang},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2025}
}