Reproduce “VINS-120K: Ultra High-Resolution Image Editing with A Large-Scale Dataset” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhizhou Chen, Shanyan Guan, Zhanxin Gao, En Ci, Yanhao Ge, Wei Li, Zhenyu Zhang, Jian Yang and 1 more
Researchers from Nanjing University and vivo developed VINS-120K, the first large-scale, ultra-high-resolution (UHR) image editing dataset, along with a high-frequency-aware post-adaptation strategy to extend non-high-resolution models for direct UHR editing. This approach substantially improved texture fidelity, evidenced by a patch-FID of 9.15 on a new 4K benchmark, and produced clearer, more detailed outputs compared to existing downsample-upsample methods.
Key insights
- Directly performing edits in the UHR space, rather than downsampling and then upsampling, is crucial for preserving fine-grained details and maintaining overall visual fidelity.
- Effectively adapting Transformer-based NHR models to UHR contexts requires specific modifications to attention mechanisms and positional encodings (e.g., attention score and RoPE rescaling) to manage exponentially longer token sequences and avoid performance degradation.
- Integrating explicit frequency-focused supervision during training is vital for synthesizing perceptually realistic textures and accurately preserving high-frequency details in UHR image editing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.