- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
The paper focus on video outpainting when objects are only partially visible in the scene. The author proposes a retrieval-augmented test-time adaptation framework built on Wan2.1-VACE video generation model. It retrieves related videos, segments the reference object, and then converts sampled superpixel regions into reference tokens for appearance adaptation. And then adapts spatial and temporal attention heads in two stages using two LoRAs. The paper also introduces ObjectBench, which evaluates object appearance through DINOv2 features and motion through point trajectories.
Critical strengths
- The main idea is intuitive and well motivated. When the input shows only part of an object, retrieving related videos provides additional visual information that can help complete the missing parts.
- The qualitative results show more complete object structures and better motion consistency than the compared methods, particularly when objects are heavily truncated.
Critical weaknesses
- It is unclear how the ObjectBench is built. The paper explains how scores are calculated, but not how objects and crops are selected or how generated objects are segmented. It also does not say whether scoring includes the unchanged input regions, which could hide errors in the generated parts.
- The experiments do not fully show why the proposed designs work. The paper does not test whether its attention-head grouping is valid or whether sampling small object regions is better than using the whole object. Adding retrieval without decoupling also makes LPIPS and FVD worse than the baseline, which needs further explanation.
- Several implementation details and the processing cost are missing. Specifically, the RAG part might take much more computation and time.
Questions for authors
-
Spatial-temporal head classification lacks direct validation against alternative or random groupings. Moreover, removing decoupling worsens LPIPS and FVD relative to the baseline. This raises questions about when retrieval helps.
-
Figure 3 is difficult to connect to the method description and equations. The masking and supervision associated with “Random frame” are not shown clearly, and WarpFusion is not explicitly defined in the text.
-
The implementation details and practical costs int the RAG part are missing. Head assignment, retrieval ranking, and background propagation are only briefly described.
-
The treatment of multiple objects is unclear. The paper does not explain whether each object receives a separate reference or how reference features are associated with the correct source object.
-
ObjectBench’s construction is underspecified. The paper describes its metrics but does not explain target-object selection, crop placement, or how generated-object masks are obtained. It is also unclear whether scoring covers the full object or only the generated regions.
-
OSA randomly samples one SLIC region per adaptation step, without explaining why this should establish useful part-level correspondences. The treatment of irregular regions, including cropping, resizing, padding, and positional information, is missing.
Thank you to the authors for their responses. As most of my concerns have been addressed, I have raised my score.