Reproduce “Representation Conditioning for Generation” 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
Saeed Razavi, Xingjian Leng, Liang Zheng
RECON conditions later diffusion-transformer blocks on semantic cues from a frozen vision encoder during training, then retains its trained projection head to provide proxy cues at generation. On ImageNet 256×256 with SiT-XL/2, RECON reaches gFID 3.25 versus REPA’s 7.90 after 400K iterations without classifier-free guidance. It requires REPA-style alignment; this is an iteration comparison, not elapsed training time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.