hUwL (score 4 / conf 3)

Major Weaknesses:

• The claim of "scene-awareness without sacrificing semantic diversity" is overstated by the reported numbers. RP@3 and FID are worse compared to the MDM baseline.
• Without training with the text-scene-motion triplet, how does the model handle scene-related text prompts? For example, is the model able to generate motions properly with a prompt like "Go and sit on the chair"? This should be discussed as a potential limitation of the proposed method.

Justification Of Preliminary Recommendation:

The proposed method is well-motivated and technically sound. But the paper shouldn't overstate and should discuss the limitations properly.

<aside> 📝

without sacrificing → tradeoff가 적음을 강조하고 싶었다 / 고치겠다

limitation → 안 되는거 정리해서 주고 revision때 수정하겠다고 얘기

</aside>

Qg3t (score 4 / conf 3)

Major Weaknesses:

• Using inbetweening as the proxy task, the paper's central design decision, is never rigorously motivated. Other text-free alternatives (masked completion, trajectory prediction, motion denoising) could serve the same role. Tab. 5 shows the pipeline works, but not that inbetweening is the best choice.
• The evaluation set is assembled by matching HML3D pairs with trajectory positions sampled from TRUMANS, which may favor SceneAdapt. Evaluation on an independently collected scene set would strengthen validity.
• Missing comparison with SceneMI [20]. SceneMI directly addresses scene-aware inbetweening using scene-motion data, highly overlapping with this work's problem formulation.

<aside> 📝

다른 evaluation set에서 평가 ? : 뭘 써야 한담?

SceneMI evaluation해서 추가로 report ?

Justification은 밑에서 같이

</aside>

• New eval set ?

Alternatives

• motion denoising?
  • x0.5 → x1