Reducing error in extreme poses

Three ways of using Refine with animation, tried on the same dancing character (LOD 4, about 6 % of the original triangles). Each worked; which one works best depends on the character and the animation.

Before you start

• Tick the clips you want the LOD to hold up in, then Generate LODs. Refinement picks its frames from what this step measured.
• Select the level to optimise in the LOD table.
• Note the maximum errors in the timeline header, so you can compare. In this example, after Generate LODs: 3.07 / 8.13 (surface / silhouette).

A. Automatic: the worst frames

Frame source: Auto (top-K worst frames), K 16, Surface budget 70 %, then Refine with animation. Result: 1.85 / 3.95.

Changing K or the budget changes the result, and not always in the obvious direction — see Examples & measurements. Refining takes seconds, so trying two or three values is practical.

B. Evenly spread frames

Frame source: Uniform across clip, K 16. On this character: 2.36 / 5.79 — a clear improvement over no refinement, but less than A. On a second character with a flatter error curve, this option came out ahead (see A new character from scratch).

C. Peak by peak, with marked ranges

A more hands-on approach that puts you in charge of which poses count:

1. Set Frame source to From marked ranges.
2. In the timeline, find the highest peak and Ctrl+Shift+click it — a 3-frame range is created around the local maximum.
3. Refine with animation. The marked peak comes down; another frame becomes the highest.
4. Mark that one too and refine again. Repeat a few times, noting the header values.

Step	Range added (frames)	Surface	Silhouette
after Generate	—	3.07	8.13
1	451–453	2.12	4.28
2	870–872	1.98	4.71
3	983–985	2.32	4.16
4	368–370	2.14	4.14
5	897–899	1.87	4.27
6	471–473	1.91	3.61

• The first range did most of the work.
• The values do not fall at every step: fixing one peak can let another frame get worse. Keep an eye on the numbers and stop at a step you like.
• The Ranges list shows the error inside each range, which helps to see which peaks responded.

  

Side by side

Approach (same character, LOD 4)	Surface	Silhouette
No refinement	3.07	8.13
A. Auto, K 16	1.85	3.95
B. Uniform, K 16	2.36	5.79
C. Six marked ranges	1.91	3.61

On this character all three helped. A gave the lowest surface error, C the lowest silhouette error, at the cost of a few minutes of work. Another character may rank them differently.