Towards Coherent Image Inpainting Using Denoising Diffusion Implicit Models

Authors: Guanhua Zhang, Jiabao Ji, Yang Zhang, Mo Yu, Tommi Jaakkola, Shiyu Chang

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our experiments verify that COPAINT can outperform the existing diffusion-based methods under both objective and subjective metrics. Our experimental evaluations on Celeb A-HQ and Image Net with various shapes of the revealed region verify that COPAINT has better inpainting quality and coherence than existing diffusion-model-based approaches under both objective and subjective metrics. Quantitative results Table 1 shows the quantitative results of the proposed COPAINT-FAST, COPAINT and COPAINT-TT together with all other baselines on both Celeb A-HQ (top) and Image Net (bottom) datasets with seven mask types.
Researcher Affiliation Collaboration 1UC Santa Barbara 2MIT-IBM Watson AI Lab 3IBM Research during the project s involvement 4MIT CSAIL.
Pseudocode Yes Algorithm 1 COPAINT-TT
Open Source Code Yes The codes are available at https://github. com/UCSB-NLP-Chang/Co Paint/.
Open Datasets Yes we validate our method on two commonly used image datasets: Celeb A-HQ (Liu et al., 2014) and Image Net-1K (Russakovsky et al., 2015).
Dataset Splits Yes We use the first five images in the validation set for hyperparameter selection. The first 100 images in test sets are used for evaluation following Lugmayr et al. (2022).
Hardware Specification Yes All experiments are done on an Nvidia-V100-SXM2-32GB GPU.
Software Dependencies No The paper mentions using 'DDPM sampler' and building upon the 'REPAINT code base', but it does not provide specific version numbers for ancillary software dependencies like PyTorch, TensorFlow, or other key libraries used in their implementation.
Experiment Setup Yes For all methods, we set the number of reverse sampling steps as 250 if not specified otherwise. Specifically, we set gradient descent step number G = 2 for both COPAINT and COPAINT-TT. A time-efficient version of our method, COPAINT-FAST is further introduced with G = 1 and reverse sampling step number as 100. We adopt an adaptive learning rate as ηt = 0.02 αt for all our methods. For COPAINT-TT, we use time travel interval τ = 10 and travel frequency K = 1.