Solving Diffusion ODEs with Optimal Boundary Conditions for Better Image Super-Resolution

Authors: Yiyang Ma, Huan Yang, Wenhan Yang, Jianlong Fu, Jiaying Liu

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our method on both bicubic-SR and real-SR degradation settings. For bicubic-SR, we train a vanilla diffusion-based SR model which simply concatenates LR images with noisy images xt as the architecture proposed in SR3 (Saharia et al., 2022c). For real-SR, we apply our method to Stable SR (Wang et al., 2023), which finetunes pre-trained Stable Diffusion (Rombach et al., 2022) on real-SR data. Experiments show that the quality of SR images sampled by few-step diffusion ODE samplers with our explored BC x T significantly outperforms the quality of results sampled by existing methods owning the same architecture.
Researcher Affiliation Collaboration 1Wangxuan Institute of Computer Technology, Peking University, 2Microsoft Research, 3Pengcheng Laboratory
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code No The paper provides links to the source codes of 'compared methods' (third-party baselines) but does not provide concrete access to their own source code for the methodology described in this paper.
Open Datasets Yes For bicubic-SR, we train the model on the widely-used dataset DF2k (Agustsson & Timofte, 2017; Lim et al., 2017)... To test the performances of bicubic-SR, we use 3 different datasets containing DIV2k-test (Agustsson & Timofte, 2017), Urban100 (Huang et al., 2015), B100 (Martin et al., 2001).
Dataset Splits No The paper specifies the datasets used for training and testing, and how test data was prepared, but does not explicitly provide information about a validation split for their model's training process.
Hardware Specification Yes The total training cost is about 2000 Tesla V100 GPU hours.
Software Dependencies No The paper mentions 'torchvision (Py Torch-Contributors, 2017)' but does not provide specific version numbers for PyTorch or other key software dependencies.
Experiment Setup Yes We first train the model for 2M iterations with a batch size of 16, then train the model for another 1M iterations with a batch size of 64, ensuring the convergence of our model. We use Adam optimizer (Kingma & Ba, 2015) during the whole training process and use mixed precision to accelerate training. In practice, the R and K are set to 300 and 1,000 respectively for both bicubic-SR and real-SR.