- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
The paper proposes CueInsert, a training-free neuro-symbolic framework for intent-driven video object insertion that explicitly separates planning from rendering. Given a background video, a reference object, and a high-level relational intent (e.g., follow, approach, avoid, or parallel), the system generates candidate insertion plans over scale, placement, orientation, and trajectory with an MLLM, verifies and repairs them using symbolic spatiotemporal predicates, and renders the selected plan through coarse-composite and mask-conditioned diffusion sampling. The authors also introduce DAVIS-Ins, a curated benchmark containing 576 intent-conditioned insertion instances derived from nine DAVIS videos and 72 ordered background–object pairs, and Insertion Action Coverage (IAC), a residual-motion metric quantifying the temporal coverage of inserted-object actions beyond global translation. CueInsert achieves the highest average score and leads on several metrics among the evaluated baselines. The paper additionally evaluates pre-rendering relation satisfaction, reports a before/after symbolic-repair ablation, and presents a user study.
Critical strengths
-
The roles of the MLLM (semantic proposals and scale-plausibility judgments) and symbolic components (verification and repair) are explicitly delineated.
-
DAVIS-Ins and IAC provide a structured evaluation setting for intent-driven insertion and could benefit the community beyond this work.
-
The framework supports different relational behaviors for the same object and background without additional model training or inversion, providing a practical interface for controllable video insertion.
Critical weaknesses
-
Reliance on external components and insufficiently characterized failure propagation. The framework heavily relies on MLLM planning and judging, object detection/tracking, and monocular depth. While symbolic repair addresses violations in proposed trajectories, the main paper does not sufficiently characterize robustness when the underlying anchor observations, depth estimates, or semantic judgments are themselves incorrect. Errors in these components may propagate through both planning and rendering, and improved predicate satisfaction with respect to an erroneous anchor track would not necessarily imply correct interaction in the scene.
-
Limited benchmark diversity. DAVIS-Ins contains 576 intent-conditioned insertion instances, but these are derived from only nine side-view DAVIS videos and 72 ordered background–object pairs (Sec. 3, pp. 6, lines 302–308). The number of instances therefore does not translate into comparable diversity of underlying scenes and viewpoints. Although the benchmark is carefully curated, the main paper provides limited evidence of generalization to broader scene geometries, substantial camera motion, changing viewpoints, and complex occlusions.
-
Insufficiently established methodological novelty and incomplete controlled comparisons. Place-it-R1 [1] already describes MLLM scene and interaction reasoning, automatic insertion trajectories, and reasoning-guided diffusion for video object insertion, yet it is neither cited nor compared in the provided manuscript. VOID (ECCV 2026) [2] also uses VLM-based causal reasoning to identify interaction-affected regions and translate them into mask guidance for video diffusion. Although VOID addresses object removal rather than insertion, it further demonstrates that reasoning about object interactions to guide video editing is not new. Thus, neither MLLM planning before generation nor interaction-aware reasoning alone establishes a novel paradigm. CueInsert’s explicit relational verification and trajectory repair are potential distinctions, but their incremental contribution remains insufficiently demonstrated through controlled comparisons. I am therefore unconvinced that the paper establishes substantial methodological novelty beyond combining existing components, rather than merely concerned about missing citations.
-
Limited sensitivity and efficiency analysis. The main paper does not report systematic sensitivity experiments for anchor-tracking noise, depth errors, or MLLM mispredictions. It also does not report runtime/cost profiling as candidate count, repair iterations, and MLLM calls vary. These analyses would help establish both the robustness and practical overhead of the proposed pipeline.
-
Inconsistent aggregate planning statistics. Sec. 4.3 (pp. 8, lines 422–423) reports Feasible-1, Feasible-2, and average valid plans per case as 83.97%, 68.22%, and 2.38, respectively. However, the Overall row of Table 3 (pp. 9) reports 82.72%, 68.38%, and 2.45. The manuscript should reconcile these values and clarify the evaluation population and aggregation procedure. This inconsistency does not by itself establish that the experiments are invalid, but it makes the aggregate planning results difficult to interpret confidently.
-
Unclear verification criteria. The title emphasizes “verifiable plans,” but the main paper does not clearly explain how a plan is actually judged to pass or fail verification. Sec. 2.2 introduces relation predicates based on relative positions and motion, but does not give the concrete checks and thresholds for each relation. For example, what exactly must a trajectory satisfy to count as “follow” or “approach”? Without these details, it is difficult to reproduce the verification results.
[1] Place-it-R1: Unlocking Environment-aware Reasoning Potential of MLLM for Video Object Insertion. arXiv:2603.06140. [2] VOID: Video Object and Interaction Deletion. ECCV 2026. arXiv:2604.02296.
Questions for authors
1.How robust is the pipeline to anchor-tracking noise, depth errors, and MLLM mispredictions? Can you provide controlled sensitivity experiments and planning runtime/cost measurements as candidate count, repair iterations, and MLLM calls vary? In particular, how are incorrect anchor observations distinguished from violations in an otherwise correctly observed candidate trajectory?
2.Can you provide additional evaluation under substantial camera motion, changing viewpoints, and complex occlusions? Table 3 evaluates relation satisfaction before rendering, while Table 4 reports subjective intent-consistency ratings, including 4.41/5 for CueInsert. Since IAC does not directly measure relational correctness, can you additionally report per-intent relation satisfaction on the final rendered videos, including how often predicate-certified plans remain relation-consistent after rendering?
3.How does CueInsert differ methodologically from Place-it-R1’s MLLM reasoning and automatic insertion planning? Could comparisons with MLLM-only and simple anchor-relative rule-based planners under the same renderer isolate the contribution of symbolic verification and repair? A direct Place-it-R1 comparison, if feasible, would also be informative, with differences in backbone and training requirements clearly documented.
4.Can you reconcile the aggregate statistics in Sec. 4.3 with Table 3? Which values are correct, and do the text and table use the same set of evaluated cases and the same aggregation procedure?
5.How exactly do you verify a plan? Please state the checks and thresholds used to decide whether it passes or fails each requested relation, and provide an example of a passing and a failing trajectory. If these details are in the supplement, please indicate where.