Reproduce “DAGS: Disentangled Appearance-and-Geometry Steering of a Frozen Image DiT for Temporally Stabilized Generative Rendering” 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
Karthik Mohan Kumar, Damian Andrysiak, Pedro Antonio Pena, Kunal Tyagi, Rama Harihara
DAGS addresses noisy one-sample path-traced rendering by steering a frozen image-generation model with separate geometry and appearance signals, then applying temporal stabilization outside it. Across nine scenes, per-frame reconstruction reaches 29.79 dB versus OIDN’s 21.23 dB, and adding both temporal mechanisms lowers consecutive-frame LPIPS from 0.0115 to 0.0053. Comparisons are not compute-matched, and DAGS is not real-time.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.