Reproduce “General Image-to-Image Translation with One-Shot Image Guidance” 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
Bin Cheng, Zuhao Liu, Yunbo Peng, Yue Lin
The Visual Concept Translator (VCT) framework enables general image-to-image translation by extracting complex visual concepts from a single reference image while preserving content from a source image. This method achieved a LPIPS score of 0.35 and a CLIPScore of 0.66, outperforming state-of-the-art methods in quantitative style transfer tasks.
Key insights
- Complex visual concepts are more intuitively and accurately conveyed through a single reference image than through text descriptions.
- Content preservation during I2I translation can be robustly achieved by optimizing unique source embeddings for each denoising timestep (PTI) and extracting attention maps from a content-matching branch.
- Fusing multiple conditional guidance signals directly in the ϵ (noise prediction) space allows for fine-grained control over the influence of different visual attributes during image generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.