Animation-aware refinement

A LOD built from the bind pose can fail in poses the bind pose never shows. The deviation timeline measures this over your animations, and Refine with animation rebuilds the LODs taking selected poses into account.

Deviation timeline

• The curves cover all ticked clips, one after another. Frame numbers are counted across them.
• The header shows the max Surface error and max Silhouette error over the visible part of the timeline — zoom out fully (fit button) to read the maximum over everything.
• The vertical axis scales itself to the highest value, so the tallest peak always reaches the top. Compare the numbers in the header, not the height of the curve, when judging an improvement.
• Buttons next to the title: zoom out, zoom in, fit.
• Click (or drag in the strip under the curves) to move the playhead; the viewports follow.

Frame ranges

Ranges mark frames you care about — the worst moments of the animation, or poses that must look right. They are listed under Ranges with their own surface and silhouette error, and they are what the From marked ranges frame source refines for.

On the timeline	Result
Drag	Create a range over the dragged frames
Drag a range edge	Resize it
Drag inside a range	Move it
Ctrl+click	One-frame range
Ctrl+Shift+click	Find the highest point within ±5 frames and create a 3-frame range around it

Refine with animation

Refine with animation rebuilds all LOD levels from the original mesh, this time taking a set of K animation poses into account. It optimises for the level selected in the LOD table.

Frame source

Frame source	Which poses are used	Tends to suit
Auto (top-K worst frames)	The K frames with the highest error, measured when the LODs were generated.	Animations whose error curve has clear, isolated peaks.
From marked ranges	Frames inside the ranges you marked.	When you know which moments matter, or want to work peak by peak.
Uniform across clip	K frames spread evenly over the evaluated clips.	Error curves without pronounced peaks.

These are tendencies observed on sample characters, not rules — see Examples & measurements. Trying more than one is cheap.

K poses and Surface budget

• K poses — how many poses are evaluated (1–256; pick from the list or type a number). More poses is not automatically better: on a sample character K = 16 did better than K = 32 or 64.
• Surface budget (%) — how the K poses are split between the two error measures: 0 = all chosen for silhouette error, 50 = balanced, 100 = all for surface error.

How refinement behaves

• It starts from what Generate measured. The Auto frame source picks its frames from the error recorded by the last Generate LODs. If you add or tick a clip afterwards, run Generate LODs again first; otherwise the new clip's frames are not known to it.
• It replaces the LOD set. Each refine rebuilds all levels from the original mesh; the result does not build on the previous refine. Pressing it twice with the same settings gives the same LODs.
• To explore, change a setting — frame source, K, budget, or the marked ranges — and refine again.
• A refine of a character with around 20 000 triangles took about 12 seconds on a desktop PC, even with K = 64.