SkipDiff: Adaptive Skip Diffusion Model for High-Fidelity Perceptual Image Super-resolution

Authors: Xiaotong Luo, Yuan Xie, Yanyun Qu, Yun Fu

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

Reproducibility Variable Result LLM Response
Research Type Experimental We experimentally observed that the perceptual quality and distortion change in an opposite direction with the increase of sampling steps. ... Extensive experimental results show that our Skip Diff achieves superior perceptual quality with plausible reconstruction accuracy and a faster sampling speed.
Researcher Affiliation Academia Xiaotong Luo1, Yuan Xie2*, Yanyun Qu1*, Yun Fu3 1School of Informatics, Xiamen University, Fujian, China 2School of Computer Science and Technology, East China Normal University, Shanghai, China 3Northeastern University
Pseudocode No The detailed procedure is provided in supplementary materials. The main paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any concrete access information (e.g., a URL or explicit statement of code release) for the source code of its methodology.
Open Datasets Yes For face SR, we train the models at 16 × 16 → 128 × 128 on Flickr-Faces-HQ (FFHQ) dataset... For natural SR, we train at 64 × 64 → 256 × 256 on the high-diversity Image Net 1K (Russakovsky et al. 2015) dataset
Dataset Splits No The paper mentions training and evaluation datasets, but does not provide specific details on training/test/validation splits (e.g., percentages or exact counts for each split).
Hardware Specification Yes The entire Skip Diff is trained and evaluated on 1 NVIDIA Tesla V100 cards.
Software Dependencies No The paper mentions using DDPM and PPO algorithm, and refers to configurations from other papers (e.g., Saharia et al. 2022), but does not provide specific version numbers for software dependencies like libraries or frameworks.
Experiment Setup Yes Following (Li et al. 2022), we set the total diffusion steps T as 100... We train SPN for 50k iterations with a learning rate of 3e-4. The hyperparameters for optimizing SPN are set like (Schulman et al. 2017; Wang et al. 2020): the clipping parameter ϵ = 0.2, γ = 0.7, c1 = 0.5, c2 = 0.01 and λ = 1.