Reproduce “IRIS: Implicit Rendering Matters for Pose-Free Novel View Synthesis” 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

Wenyu Li, Sidun Liu, Peng Qiao, Yong Dou, Tongrui Hu

IRIS presents a fully self-supervised approach to novel-view synthesis from unordered, unposed images, combining an implicit neural representation with geometry-aware point querying and learned ray aggregation. It achieves PSNR/SSIM/LPIPS scores of 31.272/0.920/0.117 on Re10K and remains effective with randomly ordered views, while outperforming explicit Gaussian-based alternatives in rendering quality but requiring substantially higher inference time.

Key insights

  • Geometry can be injected into an implicit representation through camera-conditioned projection and multi-view feature aggregation, without maintaining an explicit point, voxel, or Gaussian scene model.
  • The learned ray renderer is important: substituting standard volume rendering reduces Re10K performance from 31.272 to 28.058 PSNR and from 0.920 to 0.868 SSIM, while a 3DGS-style substitute fails to train.
  • A multi-view pretrained backbone provides a useful initialization, but fine-tuning is necessary; simply mixing more datasets does not consistently improve domain generalization.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.