1. response should explain the choice of MDM as the backbone and discuss how the introduced ideas are applicable to more recent text-to-motion models.

2. 이거는 찾아보고

3. Trumans has higher semantic coherence with respect to humanise triplets. These differences could contribute to their poor semantic performance. Training a direct adapter trained with the same data, preservation losses and training budget, reporting both semantic and collision metrics, would provide stronger attribution. → Trumans는 text 가 없고 오로지 action이다

4. TesMo는 근본적으로 T-M-S triplet 이 필요하다. . LINGO → text가 거의 없다싶다

6. Because during stage 2 it already learned how to perform inbetweening, which makes the scene aware inbetweening objective during stage 3 inject only “geometry information” to perform better inbetweening. This geometric information is still injected when pertorming text-to-motion as well

7. CaKey recieves keyframes and as stated in L274 we use Null embedding instead of text. We did not compare with SceneMI as our goal is not to perform scene-aware inbetweening, but just utilizing as a proxy task to achieve scene-aware text to motion generation.

#TODO: HumanML3D reference score 측정해보기

Draft (261101)

Q1. response should explain the choice of MDM as the backbone and discuss how the introduced ideas are applicable to more recent text-to-motion models.

[Backbone] We use MDM as MDM is an established reference for evaluating adaptation of a pretrained motion prior. Its core mechanisms remain in recent models: DART retains DDPM denoising, clean-data prediction and text dropout for classifier-free guidance; MotionStreamer uses text/history-conditioned diffusion to predict latent noise. Both retain conditional denoising across different representations. Thus adding learned conditioning is compatible with this model family, with interfaces adjusted to each architecture.