- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes Navigational Compositional Guidance (NCG), a video generation model that provides dynamic camera trajectory guidance to help users achieve professional compositions. Key contributions include a synthetic data generation pipeline (using outpainting and virtual cameras), a two-stage training paradigm (SFT and GRPO), and a direction-aware metric (DAPE). The method is validated through automated metrics and human evaluations.
Critical strengths
-
Introducing GRPO to video composition is effective. It successfully addresses the "one-to-many" mapping issue and the lack of a stopping mechanism, empowering the model to autonomously seek aesthetic optima and achieve terminal settling.
-
The authors solve the extreme scarcity of expert kinetic data by "reverse-engineering" trajectories. Using image outpainting and virtual camera projection on high-quality static photos generates realistic camera motion videos without expensive manual data collection.
Critical weaknesses
-
The evaluation relies heavily on metrics directly optimized during training. The model is explicitly trained to maximize the PQA reward via GRPO, yet PQA is used as a primary comparative metric in Table 2, violating independent evaluation principles. Similarly, the DAPE metric is trained using temporally reversed videos as negative samples and then evaluated on its ability to detect reversed videos. Even if the scenes do not overlap, testing on the exact artificial task seen during training evaluates a learned shortcut rather than true aesthetic direction awareness.
-
As a generative model, NCG must "hallucinate" out-of-frame content when starting from a truncated view (e.g., half a face). In real-world photographic guidance, the physical scene will likely not match the AI's hallucinated pixels. This fundamental disconnect between generative hallucination and physical reality severely undermines the practical utility of the proposed "guidance."
Questions for authors
-
Since DAPE was explicitly trained to penalize temporally reversed videos, testing its accuracy on reversed videos seems circular. How does DAPE perform on independent, non-reversed video pairs of varying compositional quality?
-
Regarding the out-of-frame hallucination issue, how do you propose to resolve the practical misalignment when the user moves the physical camera, but the real-world scene does not match the AI's hallucinated pixels?
2 comments
To GtPS. This response addresses Critical Weakness 2 and Question 2 under “Questions for Authors.”
To GtPS. This response addresses Critical Weakness 2 and Question 2 under “Questions for Authors.”
We acknowledge that generating a camera trajectory from a single image may introduce synthesized out-of-frame content, even when that content appears realistic and plausible. Such content is not guaranteed to match the physical scene, and this is a limitation of generative approach. Some recent state-of-the-art composition optimization methods, such as PPC and PhotoFramer, also introduce out-of-frame hallucinations to some extent.
Our objective is to provide a visual reference for composition adjustment, emphasizing aspects such as subject placement, relative scale, and spatial balance. For short-range adjustments, we expect some discrepancies in incidental scene details to be tolerable when the suggested compositional change remains applicable to the real scene.
Our formulation assumes a static scene, which reduces complications arising from independently moving objects but does not eliminate uncertainty about unseen content. Accordingly, we intend the generated video to illustrate a possible compositional adjustment, rather than serve as a faithful prediction of the scene that the user will observe after moving the camera.
Therefore, although hallucinations may have some impact, we believe that they neither create a fundamental disconnect from physical reality nor substantially undermine the practical utility of the proposed guidance.
To GtPS. This response addresses Critical Weakness 1 and Question 1 under “Questions for Authors.”
To GtPS. This response addresses Critical Weakness 1 and Question 1 under “Questions for Authors.”
We thank the reviewer for highlighting the distinction between reward-aligned evaluation and independent evidence, as well as the limitations of reversal-based testing.
1. Evaluation beyond the optimized PQA reward.
We agree that PQA scores are not independent of our GRPO objective and should be interpreted as reward-aligned measurements. However, our evaluation also includes VBench 2.0, DeQA, Q-Align, and human assessments, which are not used as the GRPO reward.
These evaluations provide complementary evidence: VBench 2.0 assesses image quality, aesthetic quality, motion smoothness, and dynamic degree, while DeQA and Q-Align provide additional image-based assessments. These dimensions are related to, but not interchangeable with, PQA’s VQ, MQ, and CA scores.
2. DAPE training includes negatives beyond temporal reversal.
Temporal reversals are one type of negative sample, rather than the sole source of supervision for DAPE. Training also includes inferior-endpoint samples, as described in Appendix C, Table 9. Our training design follows the PQA framework and adds reversal-based supervision to encourage sensitivity to the direction of compositional change.
The motivation is documented in the newly added Appendix D.4: inspections of individual video pairs and the reported aggregate results show that PQA CA scores change little when frame order is reversed. This motivates explicit directional supervision.