Reproduce “V-Co: A Closer Look at Visual Representation Alignment via Co-Denoising” 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
Han Lin, Xichen Pan, Zun Wang, Yue Zhang, Chu Wang, Jaemin Cho, Mohit Bansal
This research systematically investigates visual co-denoising in pixel-space diffusion models, developing V-Co, a structured recipe combining a dual-stream architecture, structural masking, a perceptual-drifting hybrid loss, and RMS-based feature rescaling. V-Co-B/16 achieves an FID of 2.35 on ImageNet-256x256, matching larger models with fewer parameters and improved computational efficiency.
Key insights
- A fully dual-stream architecture that processes pixel and semantic features with separate components but joint self-attention significantly enhances co-denoising effectiveness.
- Defining unconditional predictions for classifier-free guidance through structural masking (specifically, semantic-to-pixel masking) outperforms input-level dropout strategies.
- Combining instance-level perceptual loss with distribution-level drifting loss into a hybrid objective provides superior semantic alignment compared to either loss alone.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.