Use LINEAR interpolation when sampling anims

[pjoe]

Not sure why this ever used STEP :)

With LINEAR you can get decent result with sampling every few frames (e.g. 3-4 depending on your animation), instead of the chunky result you will get with STEP

[pjoe]

Sample file, with sampling every 4 frames:

anim_sampling.zip

Tried capturing the difference in gifs

STEP:

LINEAR:

[UX3D-becher]

Although the glTF specification to my knowledge would allow single keyframes with linear interpolation, some engines expect at least two. In order to avoid breaking things for these probably very common cases, we have to resort to 'STEP' interpolation.
A user that does not want STEP interpolation at the moment can bake the animation manually.
If we want to automate the baking process, animations with only a single control point have to explicitly be handled

[pjoe]

@UX3D-becher OK, I understand the issue, but for end user this really should be handled automatically. I'll try to see how difficult it is to handle the case with only 1 keyframe.

Having to manually bake anims is a complication of the workflow, and also to some extend makes the export process less 'non-destructive'.

[pjoe]

@UX3D-becher something like this should do it 92c50f7

[emackey]

@UX3D-becher I find that reasoning a little odd. Wouldn't it be better to file bug reports against engines that can't handle an individual linear keyframe?

I guess having a workaround here doesn't hurt much, but thus far the pattern has been to insist on filing issues and fixing problems rather than bending the format spec to accommodate bugs.

[emackey]

I'm in favor of this, with or without the one-keyframe workaround.

[emackey]

Fixes #369

[donmccurdy]

Agreed that LINEAR interpolation should be used for baking whenever possible. Engines are not rendering at 24-30 fps, which the keyframes usually are. Whether single-keyframe tracks should use STEP or LINEAR I'm not sure.

[donmccurdy]

Hm, but this will go to STEP for all channels, if any channel has a single keyframe? I might prefer LINEAR in that case, unless it's easy enough to specify per-channel.

[emackey]

Ouch. Would definitely prefer LINEAR in that case.

[pjoe]

I am all for always using LINEAR and would consider it a weird edge case / bug if engines don't handle it correctly. So I can easily revert last commit if we agree on this? Or try to add further detection of single channel (won't there always be 3 or 4 together?) Anyway this PR improves behavior in most common cases (more then 1 key), so can also be merged as is and we can further tackle the single key/STEP issue in separate PR

[pjoe]

After thinking/testing this some more:

I believe the best option is to use LINEAR if any of the channels in a sampler has more than 1 keyframe,
since loc/scale/rot will get each their sampler and glTF will batch together each subcomponent, i.e. glTF doesn't support having different interpolation (or number of keys) for e.g. loc.x and loc.y anyway.

If anyone can come up with a case where this would break - speak up 😄

Otherwise I think it is good to go now

[emackey]

Any objections @UX3D-becher @donmccurdy ?

[donmccurdy]

Sounds good to me. I could imagine CUBICSPLINE causing problems with a single keyframe, but STEP and LINEAR seem completely reasonable. If there's a bug in e.g. threejs we can certainly fix those.