Anti-aliasing slow in MR?

[voxel]

We've been rendering tests in MR for the past month or so and found MR a fairly fast and reliable renderer except when you need to increase the anti-aliasing settings which then turns 5-10 minute renders into hour long frames (Ideal: 0,3 with 0.025 contrast ratios - Reality: 0,2 with 0.04 contrast ratios). Mental Ray seems to need high anti-aliasing settings to render bevels and thin geometry without noise/flickering - at least compared to Maya's default render. Any hints?

BTW We aren't using motion blur or displacement, just vanilla FG (which is fast).

[bart]

Are these polys, subds?

Approximation settings?

Is smoothing on?

Where are these placed with respect to world origin?

Are you translating to mi files first?

[voxel]

Oops. MR for Maya 6.5 - 100% polygons. A large exterior set. No approximation. All surfaces hand-smoothed (i.e a sphere is constructed out of a smoothed cube and baked - senior artist preferred this). The default production settings: 0,2 - 0.1 contrast ratio still has objectionable noise along bevels and thin strips of geometry - BUT the render times are tiny (with FG or not). Adding fast SSS doesn't increase the render times much either. However, lowering the contrast ratios produces a serious performance hit.

I understand 0 means (1 sample per pixel), 1 (4 samples per pixel), 2 (16 samples per pixel). We've isolated the rendering times to the anti-aliasing settings after playing with the FG settings and then removing it altogether.

P.S I meant 0.025 and 0.04 contrast ratios in the earlier message.


Are these polys, subds?

Approximation settings?

Is smoothing on?

Where are these placed with respect to world origin?

Are you translating to mi files first?

[bart]

Is your material at the thin geometry sections highly reflective?

Do you get the noise when turning off FG?

If so, check your world scene origin with respect to where your camera is and the problem noise areas.

Are you exporting to mi file from Alias, and then rendering in standalone?
Using binary, or ascii?

[bart]

Are you looking at samples diagnostics, as you adjust contrast?

Checking the areas where noise is located?

[voxel]

Yes, we looked at the sample diagnostics. We even tried splitting and triangulating the thin edge geometry.

The noise is geometry/edge aliasing on a standard lambert material. Using (0,3) and (0.04) gets rid of it mostly, but the render times turn from minutes to hour plus. I'll check where the geometry is with respect to the origin - I never considered precision issues.

Also, I tuned the BSP settings of a personal scene of mine and sped it up by 6x! Maybe optimizing the BSP would help - it is a rather large exterior set, but would AA performance be affected by it?

P.S We use a multi-frame FG cache when we enable FG (but for tuning AA - we remove it). FG is fast and relatively noise-free.