Reproduce “Diffusable Latents from Structure-Agnostic Distillation” 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
Adrien Ramanana Rahary, Nicolas Dufour, Patrick Pérez, David Picard
Method distills teacher features into image latents without requiring shared spatial grids, addressing the mismatch between position-wise alignment and flat token sequences or text teachers. It pools latent tokens into an image descriptor for Pool-Align, or compares relations among images. On 256×256 ImageNet, Pool-Align reaches gFID 5.77 versus VF’s 6.04, a small single-run margin; gridless and cross-modal tests are limited.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.