Image Inpainting via Iteratively Decoupled Probabilistic Modeling

Authors: Wenbo Li, Xin Yu, Kun Zhou, Yibing Song, Zhe Lin

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental On multiple benchmarks, we achieve new state-of-the-art performance. Our code and models will be publicly available.
Researcher Affiliation Collaboration 1Huawei Noah s Ark Lab, 2HKU, 3CUHK (SZ), 4Alibaba DAMO Academy, 5Adobe Research
Pseudocode No The paper describes the pixel spread scheme using numbered steps and mathematical formulas but does not present it in a formal pseudocode block or algorithm box.
Open Source Code Yes Our code and models will be publicly available.
Open Datasets Yes We train our models at 512 512 resolution on Places2 Zhou et al. (2017) and Celeb A-HQ Karras et al. (2018) in order to adequately assess the proposed method. Places2 is a large-scale dataset with nearly 8 million training images in various scene categories. Additionally, 36,500 images make up the validation split. For Celeb A-HQ, we employ 24,183 and 2,993 images, respectively, to train and test our models.
Dataset Splits Yes Places2 is a large-scale dataset with nearly 8 million training images in various scene categories. Additionally, 36,500 images make up the validation split. For Celeb A-HQ, we employ 24,183 and 2,993 images, respectively, to train and test our models.
Hardware Specification Yes We train our models for 20M images on Places2 and Celeb A-HQ using 8 NVIDIA A100 GPUs.
Software Dependencies No The paper mentions using Adam Kingma & Ba (2015) optimizer, but does not provide specific version numbers for software dependencies like Python, PyTorch, or CUDA.
Experiment Setup Yes The batch size is 32, and the learning rate is 1 10 3. We employ an Adam Kingma & Ba (2015) optimizer with β1 = 0 and β2 = 0.99. We empirically set α = 0.01 in Eq. (5) based on experimental results. During training, our model undergoes the entire pipeline with two iterations to enhance efficiency. However, during testing, the model iterates four times to achieve improved restoration results.