FlexiAct: Towards Flexible Action Control in Heterogeneous Scenarios

The paper presents a clear problem statement, proposes a logical two-component solution addressing specific challenges (spatial adaptation, action extraction), and justifies methodological choices based on observations (e.g., denoising timestep attention). The argument flows logically from problem to solution, methodology, results, and discussion.

Strengths: Detailed description of the base model (CogVideoX-I2V) and proposed components (RefAdapter, FAE)., Explicit details on training procedures (two-stage training, dataset used, training steps, learning rate, optimizer, batch size, parameter count for RefAdapter)., Definition and application of specific automatic evaluation metrics (Text Similarity, Motion Fidelity, Temporal Consistency, Appearance Consistency)., Description of human evaluation methodology (number of raters, comparisons)., Inclusion of ablation studies to demonstrate the necessity and effectiveness of each component.
Weaknesses: Specifics of the 'learnable embeddings' used in FAE are not fully detailed., The composition or acquisition process of the evaluation dataset is described, but it's based on existing data [Ju et al. 2024].

The claims are strongly supported by both quantitative results (Table 1) showing superior performance across multiple metrics and extensive qualitative results (Figures 1, 4-11) illustrating the method's capability in diverse scenarios and subjects, as well as the impact of its components (ablation studies).

The paper introduces a novel framework (FlexiAct) combining two new components (RefAdapter and FAE) to address the specific challenges of action transfer in heterogeneous scenarios. The concept of Frequency-aware Action Extraction based on denoising timestep attention dynamics appears original.

The ability to flexibly transfer actions across diverse subjects and spatial structures without strict constraints or per-video manual effort has significant potential impact on creative fields like animation, games, and film, making video customization more accessible and efficient.

Strengths: Key concepts (FlexiAct, RefAdapter, FAE) are clearly introduced and described., Methodology is explained in sufficient detail., Evaluation metrics and procedures are well-defined., Results are presented clearly in tables and figures., Academic terminology is used appropriately.
Areas for Improvement: None