Reproduce “Generative Refinement Networks for Visual Synthesis” 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
Jian Han, Jinlai Liu, Jiahuan Wang, Bingyue Peng, Zehuan Yuan
Generative Refinement Networks (GRN) from ByteDance present a visual tokenizer, Hierarchical Binary Quantization (HBQ), for near-lossless discrete compression, coupled with a global refinement mechanism to mitigate error accumulation and a complexity-aware sampling strategy for adaptive inference steps. The framework achieves an FID of 1.81 for class-conditional image generation on ImageNet 256x256 and demonstrates superior performance for text-to-image and text-to-video tasks compared to models of similar scale.
Key insights
- Hierarchical Binary Quantization (HBQ) allows for high-fidelity, near-lossless discrete representation of visual data by recursively halving the quantization error with each binary decision.
- A global generative refinement mechanism, inspired by human drawing, is crucial for mitigating error accumulation in autoregressive models by enabling iterative correction of previously generated content.
- Dynamically adjusting inference steps based on content complexity through entropy-guided sampling significantly improves computational efficiency without sacrificing output quality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.