Reproduce “Recurrent Dynamic Range Extension” 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
Sebastian Dille, Keru Fu, S. Mahdi H. Miangoleh, Yağız Aksoy
Recurrent Dynamic Range Extension reconstructs clipped highlights from a single LDR or RAW image by learning a bounded one-EV expansion and applying the same network repeatedly. The method combines RAW-based training for diverse image structure with recurrent HDR supervision, achieving strong zero-shot results on SI-HDR and competitive perceptual quality on BtP-HDR while recovering substantially larger luminance ranges than its single-step variant.
Key insights
- Breaking HDR reconstruction into bounded one-EV steps enables direct learning in linear radiometric space without logarithmic or tone-mapped targets.
- RAW-based learning supplies broad natural-image priors, whereas HDR supervision is needed for recovering long-tailed luminance and extending the range across multiple recurrent steps.
- The method can infer plausible structure in clipped highlights, but fully saturated regions remain fundamentally ambiguous and may produce incorrect textures or bright spots.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.