Reproduce “SpeeDiff: Scalable Pixel-Anchored End-to-End Latent Diffusion Model” 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
Bingliang Zhang, Wenda Chu, Yizhuo Li, Linjie Yang, Yisong Yue, Katherine L. Bouman, Yang Song, Qiushan Guo
SpeeDiff jointly trains an image encoder and diffusion generator, adding Tweedie Pixel Reconstruction (TPR) to decode the generator’s clean-latent estimate and compare it with the input pixels. On class-conditioned ImageNet 256×256 without guidance, TPR lowers gFID from 33.95 to 5.79 after 80 epochs. The result supports the pixel anchor in this setup, not attribution of later gains to TPR alone.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.