PostEdit: Posterior Sampling for Efficient Zero-Shot Image Editing

Authors: Feng Tian, Yixuan Li, Yichao Yan, Shanyan Guan, Yanhao Ge, Xiaokang Yang

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results indicate that the proposed Post Edit achieves state-of-the-art editing performance while accurately preserving unedited regions. Furthermore, the method is both inversion- and training-free, necessitating approximately 1.5 seconds and 18 GB of GPU memory to generate high-quality results.
Researcher Affiliation Collaboration 1Mo E Key Lab of Artificial Intelligence, AI Institute, Shanghai Jiao Tong University 2vivo Mobile Communication Co., Ltd EMAIL EMAIL
Pseudocode Yes Algorithm 1 Posterior Sampling for Image Editing
Open Source Code Yes Code: https://github.com/TFNTF/Post Edit.
Open Datasets Yes To ensure a fair comparison, all experiments were conducted on the PIE-Bench dataset Ju et al. (2024) using the same parameter settings specified in Appendix A.2 and a single A100 GPU to evaluate both image quality and inference efficiency. The PIE-Bench dataset comprises 700 images with 10 types of editing, where each image is paired with a source prompt and a target prompt.
Dataset Splits Yes The PIE-Bench dataset comprises 700 images with 10 types of editing, where each image is paired with a source prompt and a target prompt. In our experiments, the resolution of all test images was set to 512 × 512. For the reconstruction experiments, we set the initial and target prompts to be identical across all test runs.
Hardware Specification Yes Additionally, the method is both inversion- and training-free, necessitating approximately 1.5 seconds and 18 GB of GPU memory to generate high-quality results.
Software Dependencies No The paper mentions specific models and solvers like 'LCM-SD1.5' and 'Dreamshaper v7 fine-tune of Stable-Diffusion v1-5' but does not provide version numbers for general software libraries such as Python, PyTorch, or CUDA, which are typically required for replication.
Experiment Setup Yes The main hyper-parameters of the Post Edit are briefly summarized in Tab. 3. Parameters of Consistency models. cskip and cout shown in line 6 of Alg. 1 are set to 0 and 1 for most cases respectively. Hyper-parameters in Alg. 1. N is set to 5 for schedule {τi}N−1i=0 . To ensure higher efficiency and quality at the same time, zN is sampled through zN ∼ N αt z0, 1 − αt I , (18) where t is set to 501 generally following the DDPM scheduler Ho et al. (2020).