The best denoiser depends on what you are already paying

Path tracing is an average. With enough samples the average is the picture; with the one or two you can afford at interactive rates it is noise, and Rayxix spends as much code removing that noise as it does making it.

There are four ways to do it in there, switchable while it runs:

• Bilateral — a single edge-aware blur, guided by normals and depth.
• A-trous — the same idea over several passes at widening strides.
• Temporal — reproject the previous frame through the previous camera matrix, reject where the normal or depth says the reprojection is a lie, and clamp what survives into this frame's own variance.
• SVGF — the temporal pass with a variance estimate driving the spatial one.

The thing I expected to find

That SVGF wins, because it is the most sophisticated. It does win on quality, every time. What I did not expect is how completely that stops being the interesting number.

What the benchmark actually said

A deterministic flythrough, three presets, CSV out. Overhead is measured against no denoising at all:

• Performance — 1 SPP, 2 bounces, a 5.63 ms frame. SVGF costs +46%. Winner: bilateral, which buys 14 quality points for 11%.
• Mixed — 2 SPP, 4 bounces, a 16.46 ms frame. SVGF costs +16%. Winner: temporal.
• Quality — the trace dominates at around 103 ms. SVGF costs +5% and scores 98. Winner: SVGF, easily.

The denoiser costs roughly the same number of milliseconds in all three. What changes is what it is a fraction of. At one sample per pixel the trace is 5.6 ms and the denoiser is nearly half again on top; at the quality preset the trace is a hundred milliseconds and the same filter disappears into the rounding.

A fixed cost is not a fixed decision. It is a fraction, and you do not know the fraction until you know the frame.

The mixed preset is the one I would not have guessed at all. Temporal scores 90.9 against SVGF's 93.9 — three points worse — but runs at 18.16 ms against 19.04, and once both are folded into one score the cheaper one comes out ahead. Three points of quality nobody can see, for 5% of the frame that everybody can.

What I took from it

Build the switch before you build the opinion. The four denoisers exist because I could not tell in advance which one the thesis would end up recommending, and the honest way to find out was to make all four runnable from the same UI and let a CSV settle it.