[DDGI] Stability Improvement for ComputeVisibility(): Added Stabilizer for variance floor & Delta Clamp to avoid tails

[GapingPixel]

Added Stabilizer for variance floor & Delta Clamp to avoid tails

[Copilot]

Pull request overview

Improves numerical stability of DDGI visibility computation by adding a variance floor and clamping the distance delta to reduce extreme-tail behavior.

Changes:

• Adds a variance “stabilizer” floor based on mean^2.
• Clamps delta to mean * 3.0f before computing Chebyshev visibility.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

delta = min(delta, mean * 3.0f) can clamp delta to 0 when mean is 0 (or very small), which makes chebyshev become variance/(variance+0) (often 1) and can even produce 0/0 NaNs when variance is also 0. Since probe distance moments can legitimately have mean == 0 (e.g., empty/invalid samples), clamp using an epsilon (e.g., max(mean, eps)), clamp to an absolute distance, or gate the clamp so it only applies when mean is above a threshold.

Suggested change

	delta = min(delta, mean * 3.0f); //Tries to avoid tails @test if neccesary
	// Clamp delta using a safe mean to avoid forcing delta to 0 when mean is 0 (or very small)
	float meanForClamp = max(mean, 0.001f);
	delta = min(delta, meanForClamp * 3.0f); // Tries to avoid tails @test if necessary

[Copilot]

The newly added lines use tab indentation and introduce trailing whitespace, while surrounding code in this file uses spaces. Please re-indent these lines to match the file’s existing spacing to avoid noisy diffs and style inconsistencies.

[Copilot]

Comment has spelling/wording issues (e.g., "neccesary") and includes internal notes like "@test" that read as temporary. Please correct the spelling and either remove the TODO-style note or turn it into a clear, actionable comment describing when/why the clamp should be enabled.

Suggested change

	delta = min(delta, mean * 3.0f); //Tries to avoid tails @test if neccesary
	delta = min(delta, mean * 3.0f); // Clamp delta to limit tail contribution and stabilize visibility; adjust the factor if stronger tail suppression is necessary.

[sampavlovic]

Seems right, can't find any meaningful difference even in edge case tests though, could you give more details on the edge cases this is solving

[sampavlovic]

Iterating on this and getting somewhere, I get what it's been trying to do now

[sboxbot]

This PR has been merged upstream.