DreamClear: High-Capacity Real-World Image Restoration with Privacy-Safe Dataset Curation

Authors: Yuang Ai, Xiaoqiang Zhou, Huaibo Huang, Xiaotian Han, Zhengyu Chen, Quanzeng You, Hongxia Yang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our exhaustive experiments confirm Dream Clear s superior performance, underlining the efficacy of our dual strategy for real-world image restoration.
Researcher Affiliation Collaboration MAIS & NLPR, Institute of Automation, Chinese Academy of Sciences School of Artificial Intelligence, University of Chinese Academy of Sciences Byte Dance, Inc University of Science and Technology of China
Pseudocode No The paper describes methods through text and diagrams (e.g., Figure 2, Figure 3) but does not include explicit pseudocode or algorithm blocks.
Open Source Code Yes Code and models: https://github.com/shallowdream204/Dream Clear
Open Datasets Yes We adopt a combination of DIV2K [44], Flickr2K [2], LSDIR [39], DIV8K [22], and our generated dataset to train Dream Clear.
Dataset Splits Yes For synthetic benchmarks, we randomly crop 3,000 patches from the validation sets of DIV2K and LSDIR, and degrade them using the same settings as training.
Hardware Specification Yes The proposed Gen IR framework is built on SDXL [55] and trained over 5 days using 16 NVIDIA A100 GPUs. ... The training is conducted on 1024 1024 resolution images, running for 7 days on 32 NVIDIA A100 GPUs with a batch-size of 128.
Software Dependencies No The paper mentions specific models and optimizers (e.g., SDXL, Pix Art-α, LLa VA, AdamW optimizer) but does not provide specific version numbers for general software dependencies like Python, PyTorch, or CUDA.
Experiment Setup Yes The training is conducted on 1024 1024 resolution images, running for 7 days on 32 NVIDIA A100 GPUs with a batch-size of 128. The number of experts K in Eq. (2) is set to 3. For inference of Dream Clear, we adopt iDDPM [51] with 50 sampling steps, CFG guidance scale ω = 4.5.