Comparison with earlier motion in-betweening methods

[Cr7st]

Hello!

Thank you for your interesting work! I noticed that the comparisons in the paper primarily focus on recent diffusion-based methods (e.g., MDM, GMD, OmniControl) but do not include earlier MIB approaches, such as Qin et al.'s two-stage Transformer, Harvey et al.'s Robust Motion In-betweening

Could you provide insights on how CondMDI compares to these methods? What's the difference between these methods and diffusion-based MIB? Specifically, what are the key advantages of diffusion-based in-betweening compared to these earlier approaches? Does diffusion-based MIB yield better quality results?

Additionally, the paper uses metrics like FID and keyframe error but omits MPJPE (Mean Per Joint Position Error), which is common in MIB tasks. Would it be possible to report MPJPE for CondMDI and baselines? Or is there a particular reason for using these metrics instead of MPJPE?

[setarehc]

Hi,

Thank you for your thoughtful question and for your interest in our work!

Our comparisons focus primarily on recent diffusion-based approaches like MDM, GMD, and OmniControl, as these are the most relevant baselines for our paper's main goal: integrating sparse keyframe conditioning into text-conditioned motion generation. Earlier motion in-betweening (MIB) methods such as the two-stage Transformer by Qin et al. and Robust Motion In-betweening by Harvey et al. were not included because their problem setup is different from ours.

Specifically, many of these earlier MIB methods do not support text conditioning and typically rely on observing multiple consecutive input frames, often at the start of the sequence. In contrast, our setting involves randomly placed sparse keyframes and text prompts, which are not directly compatible with the assumptions of those models. For example, Robust Motion In-betweening assumes full past context and is not applicable to sparse or unordered keyframes.

As for the advantages of diffusion-based MIB: we found diffusion models to be particularly effective at producing smooth, high-quality transitions that maintain global coherence across long sequences. They also offer a natural mechanism to incorporate different types of conditioning signals such as text and keyframes (sparse or dense - in time or in space) through techniques such as classifier-free guidance or conditioning augmentation, which earlier deterministic methods typically lack.

Regarding MPJPE: that's a great point. In our work, we opted for FID and keyframe error as they are commonly reported in motion diffusion works and they reflect both global quality and faithfulness to conditioning signals. That said, we appreciate the suggestion and agree that including MPJPE as an auxiliary metric could be useful in future evaluations for completeness.

Thanks again for raising these points! Let us know if you’d like further clarifications or would be interested in a deeper technical discussion.