[Feature] Improve sky-region stability in single-step ApDepth inference

[Haruko386]

Motivation

Although ApDepth has already converted the original Marigold-style stochastic multi-step generation into a deterministic single-step depth perception framework, sky regions can still be challenging in some outdoor scenes.

In particular, large sky areas may occasionally suffer from local depth collapse, pseudo-texture artifacts, or unstable far-depth estimation. Recent improvements in ApDepth-G, including offset noise, SNR-weighted training, and latent gradient consistency, have already reduced this issue significantly. However, further improvement is still possible for the single-step ApDepth version.

Goal

This issue tracks the improvement of sky-region robustness in the single-step ApDepth pipeline while preserving its main advantage: fast deterministic one-step inference.

The goal is not to convert ApDepth back into a multi-step diffusion model, but to make the one-step version more stable in outdoor scenes with large sky regions.

Proposed Improvements

• Introduce sky-aware training constraints for outdoor scenes.
• Reduce unreliable depth-prior influence in sky regions.
• Suppress pseudo-depth structures caused by clouds, illumination changes, or sky texture.
• Preserve clear boundaries between sky and foreground objects such as buildings, trees, mountains, and horizons.
• Improve far-depth consistency while keeping single-step inference speed unchanged.

Possible Implementation Directions

• Add a sky-aware mask branch or sky-region guidance during training.

• Apply confidence gating to weaken unreliable depth priors in sky regions.

• Use sky-aware latent gradient weighting:

  • weak gradient constraint inside sky regions;
  • stronger gradient constraint near sky-object boundaries.

• Add a lightweight sky smoothness or variance suppression term.

• Keep inference deterministic and single-step.

Expected Benefits

• More stable sky depth estimation.
• Less sky collapse in outdoor scenes.
• Reduced pseudo-texture artifacts in sky areas.
• Better horizon and object-boundary preservation.
• Improved robustness without sacrificing ApDepth's single-step inference efficiency.

Notes

This issue is intended as a future enhancement for the single-step ApDepth framework. The design should remain consistent with ApDepth's original goal: accurate monocular depth estimation with fast deterministic inference.

[Haruko386]

Try this featrue 🤗