Recover sharp and detailed monocular depth

Establish methods for recovering sharp, detailed monocular depth predictions without sacrificing the broader performance of generative depth-estimation models.

Background

The paper identifies sharp and detailed depth recovery as an unresolved problem in modern generative depth estimation while reviewing Marigold V2. This issue is directly relevant to the paper’s investigation of pixel-space backbones, because fine spatial detail is one of the principal motivations for avoiding lossy latent representations.

References

Marigold V2 moves to modern DiTs, adds representation alignment against ground-truth features and a Sinkhorn-based second stage; it names the recovery of sharp, detailed depth among the open problems of the field.

— Iris-3B: Going Beyond the Latent with Pixel-Space Diffusion Training, Conversion and Fine-Tuning  (2610.09450 - Cai et al., 7 Oct 2026) in Section 2, paragraph “Generative priors for dense prediction”