InpDiffusion: Image Inpainting Localization via Conditional Diffusion Models
Authors: Kai Wang, Shaozhang Niu, Qixian Hao, Jiwei Zhang
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments across challenging datasets demonstrate that the Inp Diffusion significantly outperforms existing state-of-the-art methods in IIL tasks, while also showcasing excellent generalization capabilities and robustness. In this section, we describe the experiments conducted on five different datasets to evaluate the effectiveness of Inp Diffusion. Quantitative Evaluations. Table 1 shows the quantitative results of our proposed Inp Diffusion compared to six other state-of-the-art (SOTA) methods across four different inpainting types on the Inpaint32K dataset. Ablation Studies. We conduct an ablation study on the individual components of Inp Diffusion, with detailed results provided in Table 3. |
| Researcher Affiliation | Academia | Beijing Key Lab of Intelligent Telecommunication Software and Multimedia, Beijing University of Posts and Telecommunications, Beijing 100876, China EMAIL |
| Pseudocode | No | The paper describes the methodology using text and figures, but does not include any explicit pseudocode blocks or algorithms. |
| Open Source Code | No | The paper does not provide an explicit statement about releasing the source code for the Inp Diffusion methodology. While it links to a dataset repository, it does not link to the model's implementation. |
| Open Datasets | Yes | Our experiments are primarily conducted on a large dataset called Inpaint32K (Hao 2024), which contains a total of 32,000 inpainted images. This dataset is divided into four distinct inpainting types, with each type comprising 8,000 images: Traditional Methods-Based, CNN-Based, GAN-Based, and Diffusion Model-Based. Other datasets include an inpainting datasets (DID (Wu and Zhou 2021)), an AIGC dataset (Auto Splice (Jia et al. 2023)), a real-life dataset (IMD (Novozamsky, Mahdian, and Saic 2020)), and a traditional benchmark dataset (Nist (NIST 2016)). |
| Dataset Splits | Yes | It s important to note that DID is exclusively used as a test set, while the other datasets are divided into training and test sets with a 9:1 ratio. |
| Hardware Specification | Yes | We implement our Inp Diffusion based on Py Torch using a single NVIDIA A800 with 80GB memory for both training and inference. |
| Software Dependencies | No | The paper mentions Py Torch as the framework and Adam W as the optimizer, but does not provide specific version numbers for these or any other software dependencies. |
| Experiment Setup | Yes | For efficient training, the model undergoes a total of 150 training epochs. For optimization, the Adam W (Loshchilov and Hutter 2019) optimizer was utilized along with a batch size set to 32. To adjust the learning rate, we implemented the cosine strategy with an initial learning rate of 0.001. Notably, we set T = 10 for sampling and set SNR Shift to 2 log(6). In the loss function, we set the ratio of λ to µ as 7:3. |