PaperlayerView exact revision on arXiv

Source-linked AI summary

Weaving Visual Narratives: Agentic Image Bundle Composition Beyond Atomic Visual Matching

Rong Shan, Tianyi Xu, Congmin Zheng, Wenteng Chen, Jiachen Zhu, Junjie Wu, Teng Wang, Weiwen Liu, Changwang Zhang, Weinan Zhang, Jun Wang, Jianghao Lin

arXiv:2608.28695v1cs.CVcs.IR

TL;DR

Image search often involves finding a cohesive visual story whose images are related, rather than matching isolated snapshots. The paper introduces IBC and BundleWeaver’s incremental hyperedge discovery approach; experiments report state-of-the-art performance across all metrics.

  • Problem

    Users may seek cohesive image bundles whose members jointly satisfy a query, a need not captured by isolated image matching.

    Source: §1 · p. 1

  • Method

    BundleWeaver treats IBC as query-conditioned hyperedge discovery, incrementally expanding bundles by fulfilling missing roles and applying whole-bundle reranking.

    Source: §1 · p. 2

  • Results

    BundleWeaver achieves state-of-the-art performance across all metrics.

    Source: §5.2 · p. 7

  • Takeaways & Limitations

    Results indicate that independent text–image alignment does not capture cross-image relations in IBC.

    Source: §5.2 · p. 7

    The investigation is limited to static images in personal photo collections and has not yet extended to specialized domains or dynamic modalities.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Image retrieval has traditionally been formulated as a point-wise matching problem, where each candidate image is scored in isolation. However, this atomic paradigm fails to capture the complexity of human search intent within personal photo collections, where users often seek compact visual stories bound by structural relations rather than isolated snapshots. To address this limitation, we introduce **Image Bundle Composition (IBC)**, a novel paradigm that shifts the objective from ranking individual images to dynamically composing cohesive image bundles from a massive, unstructured photo pool. Since target bundles are not predefined, IBC presents a severe combinatorial explosion challenge and demands modeling non-decomposable joint relevance. To establish this paradigm, we construct **IBCBench**, the first IBC benchmark dataset containing 109,467 images and 667 verified queries, built via a semi-automated verification pipeline. Furthermore, we propose **BundleWeaver**, an agentic framework that reformulates IBC as query-conditioned incremental hyperedge discovery. By employing a Large Language Model to adaptively search for missing relational roles and utilizing a Vision-Language Model for whole-bundle verification, BundleWeaver effectively navigates the combinatorial space. Extensive experiments demonstrate that while state-of-the-art embedding models and static decompose-and-rerank paradigms suffer from relational blindness, BundleWeaver achieves substantial performance gains, highlighting the necessity of shifting from atomic scoring to dynamic relational composition. Our dataset and code are available.

1 Introduction

The introduction reframes retrieval as composing cohesive image bundles rather than ranking isolated images, and presents IBCBench and BundleWeaver to address the task’s search and evaluation challenges.

1 Introduction

  • Existing retrieval ranks independently relevant images—even multi-hop systems ultimately return disconnected results—while users often seek visual stories whose members jointly express temporal, event, or spatial relations (Lin et al., 2024; Shan et al., 2025; Deng et al., 2026).The introduction notes that standard benchmarks also retain the atomic assumption.Source: §1 · p. 1
  • Image Bundle Composition (IBC) shifts retrieval from ranking individual images to dynamically composing cohesive bundles from an unstructured photo pool, where arbitrary subsets create a combinatorial search space.Bundles are not predefined or indexed; the system must discover image combinations that instantiate the intended narrative or relation.Source: §1 · pp. 1–2
  • IBCBench contains 109,467 images and 667 verified queries, built by mining spatiotemporal-session windows, checking cross-image relations with a VLM, and having annotators refine queries and remove ambiguity.The pipeline addresses the impracticality of exhaustive subset annotation and the risk that unconstrained VLM discovery yields incoherent or non-unique bundles.Source: §1 · p. 2
  • BundleWeaver models IBC as query-conditioned hyperedge discovery, incrementally expanding seed images to fill missing bundle roles with an LLM agent, adaptive beam search, contextual pruning, and whole-bundle VLM reranking.The agent conditions its search direction on both the original query and the current partial bundle rather than greedily retrieving locally relevant images.Source: §1 · p. 2
  • On IBCBench, BundleWeaver achieves substantial performance gains over state-of-the-art multimodal embeddings and VLMs, which find isolated matches but struggle to compose them logically.The introduction attributes the gain to explicitly modeling and searching for cross-image compositions.Source: §1 · p. 2

2 Task Formulation

Image Bundle Composition replaces independent image ranking with query-conditioned selection of a compact, cohesive bundle whose images jointly satisfy relational or narrative intent. This formulation creates combinatorial search and non-decomposable relevance challenges.

2.1 Preliminaries: Traditional Text-to-Image Retrieval

  • Unlike IBC, traditional text-to-image retrieval scores each candidate image independently against the query and returns the highest-scoring image or top-k list.An image’s relevance is atomic and decoupled from the other images in the retrieved list.Source: §2.1 · p. 3

2.2 Image Bundle Composition (IBC)

  • IBC dynamically composes a compact image bundle that collectively fulfills a user’s relational or narrative intent, rather than returning one image or a disconnected list.The bundle size is determined dynamically, subject to a small upper bound.Source: §2.2 · p. 3
  • IBC selects a subset B of K distinct images to maximize joint relevance Φ(B, q), with K dynamically determined and bounded by Kmax.The formulation allows bundle sizes such as 2 ≤ K ≤ Kmax.Source: §2.2 · p. 3
  • The joint score Φ evaluates the bundle as a whole, capturing query-specified structural, temporal, or spatial relations across its images.It is not merely independent visual-textual alignment.Source: §2.2 · p. 3

2.3 Discussions on Challenges of IBC

  • Because candidate bundles are not predefined or indexed, IBC must search image subsets in a space that grows combinatorially with pool size, making exhaustive search intractable for typical personal collections.The paper identifies this as a fundamental computational challenge.Source: §2.3 · p. 3
  • IBC’s joint relevance is non-decomposable: no monotonic aggregation of individual image scores can approximate the bundle-level score.The paper contrasts this with aggregation operators such as sum or average.Source: §2.3 · p. 3
  • For a query about the Eiffel Tower’s day-to-night transition, two high-scoring daytime images can form a low-relevance bundle, so selecting the independent top-K images fails.This example motivates shifting from independent scoring to relational composition.Source: §2.3 · p. 3

3 Dataset Construction

IBCBench addresses the difficulty of building an Image Bundle Composition dataset with a five-stage semi-automated pipeline. It combines candidate mining, VLM verification, and expert review to produce verified, structurally constrained bundles.

Property Requirements

  • Accepted bundles must make every image indispensable, enforce a higher-order cross-image relation, minimize ambiguity and alternatives, and keep images visually distinct.These criteria are Joint Completeness, Cross-image Binding, Uniqueness, and Bounded Redundancy.Source: §3 · pp. 3–4

Candidate Mining

  • Mining retained 109,467 YFCC-100M photos and pruned spatiotemporal sliding windows of 3–5 images to 7,460 diverse candidates, bypassing the O(NK) search space.GPT-4o supplied dense captions and tags, while multimodal embeddings helped penalize redundancy; a composite heuristic scored spatiotemporal coverage and semantic richness.Source: §3 · p. 4

VLM Verification & Human Review

  • Claude-Opus-4.5 verified candidates against same-location dynamics and cross-location structures, generated queries and shared anchors, and four expert annotators reviewed them against C1–C4.Reviewers filtered ambiguous cases to ensure global uniqueness.Source: §3 · p. 4

Final Dataset Statistics

  • 667 high-quality verified queries form IBCBench; bundle sizes are 3 (24.3%), 4 (32.2%), and 5 (43.5%) images, with 52.5% same-location dynamics and 47.5% cross-location structural ties.The end-to-end acceptance rate was < 9%.Source: §3 · p. 4

4 Methodology

BundleWeaver addresses the combinatorial difficulty of Image Bundle Composition by treating it as query-conditioned hyperedge discovery and constructing bundles incrementally. Its method initializes diverse seeds, adaptively expands partial bundles, and reranks completed bundles for whole-bundle coherence.

4.1 Reformulation: Incremental Hyperedge Discovery

  • BundleWeaver incrementally constructs query-conditioned hyperedges, starting from one image and adding vertices that fulfill missing narrative or relational roles.The image pool forms the vertices of an implicit fully connected hypergraph, while coherent bundles are dynamically discovered hyperedges.Source: §4.1 · p. 4, §4 · p. 4

4.2 Diverse Seed Initialization

  • To avoid local-optimum collapse and near-duplicate starts (Zhou et al., 2026; Zhu et al., 2025), BundleWeaver selects semantically and visually diverse seeds from top-K retrievals.An LLM extracts a visual anchor and search direction from the query; dense multimodal retrieval searches the image pool, and each selected seed roots a parallel construction tree.Source: §4.2 · p. 5

4.3 Adaptive Hyperedge Expansion via Parallel Beam Search

  • Parallel beam-search branches expand partial bundles using plausible spatiotemporal neighborhoods, LLM-generated sub-queries for missing relations, and composite path scoring.The LLM conditions sub-queries on the original query and selected-image captions; scoring balances step-wise matches against alignment with the original query, with λ controlling the trade-off.Source: §4.3 · p. 5

4.4 Whole-Bundle VLM Reranking

  • Search terminates when the LLM deems a bundle complete, subject to Kmax, and a VLM scores completed candidates for structural coherence and cross-image binding.The VLM evaluates each candidate alongside the query on a 1–10 scale, and the highest-scoring hyperedge is returned.Source: §4.4 · p. 5

5 Experiments

Experiments show that BundleWeaver’s incremental composition and whole-bundle verification outperform atomic matching and static decomposition, with gains examined across components, bundle conditions, and VLM backbones. Evaluation uses per-query set-level metrics and exact match.

5.1 Experimental Setup

  • Evaluation averages per-query set-level Precision, Recall, and F1 alongside Exact Match; ranked-list baselines are truncated to the ground-truth bundle size.Under this oracle-size evaluation, predicted and ground-truth sets have identical sizes, making Precision, Recall, and F1 mathematically equivalent for those baselines.Source: §5.1 · p. 6

5.2 Main Results

  • Point-wise matching loses cross-image relations, metadata heuristics can degrade F1, and static decomposition improves Recall and Precision but retains low Exact Match.The passages attribute the limitations to independent image scoring, rigid spatiotemporal filtering, and the inability of independently retrieved sub-queries to enforce inter-image constraints.Source: §5.2 · p. 7
  • BundleWeaver achieves state-of-the-art performance across all metrics through adaptive hyperedge expansion and parallel beam search that balance precision with whole-bundle coherence.The reported comparison covers BundleWeaver and the IBCBench baselines; the supplied passages do not provide the table’s metric values.Source: Table 1 · p. 6, §5.2 · p. 7

5.3 Component Ablation

  • Ablations indicate that diverse seeding, contextual pruning, parallel beam search, and VLM reranking each support BundleWeaver’s performance.Without diverse seeding search can become trapped in local optima; removing pruning enlarges the noisy candidate space, while greedy search and embedding-only path scores perform worse. The pruning ablation still substantially outperforms vanilla GPT-4o, so pruning alone does not explain the gains.Source: §5.3 · p. 7

5.4 Breakdown Analysis

  • BundleWeaver outperforms baselines across bundle sizes and location types, and its performance drops less as bundles grow; static decomposition often misses the exact coherent bundle.All methods worsen with larger bundles. RzenEmbed struggles on cross-location bundles, while BundleWeaver leads for both location types; the passages link its robustness to modeling missing relational roles and verifying the whole bundle.Source: §5.4 · pp. 7–8, Figure 3 · p. 8

5.5 Backbone Generalizability

  • Across diverse open- and close-source VLM backbones, BundleWeaver consistently beats decompose-and-rerank, with a large Exact Match increase regardless of backbone.The passages attribute this generalization to incremental hyperedge discovery coupled with whole-bundle verification, rather than to a smarter VLM alone.Source: §5.5 · p. 8, Table 3 · p. 8

5.6 More Experiments

  • The paper directs readers to Appendices E and F for additional experiments and responses to potential questions.Source: §5.6 · p. 8

6 Related Works

Related retrieval and generation tasks add context or handle multiple images, but still rank images independently or assume the images are already supplied. IBC instead discovers a query-conditioned image set whose images must jointly satisfy relational, temporal, and narrative constraints.

6.1 Vision-Language Retrieval

6.2 Beyond Atomic Image Retrieval

  • Multi-image, lifelog, and composed-image retrieval return or refine sets of images, but evaluate retrieved items independently at the image level.Composed image retrieval uses visual references, while lifelog retrieval searches personal photo streams (Baldrati et al., 2023).Source: §6.2 · p. 9
  • Visual storytelling generates coherent text from images already provided, whereas IBC starts with a text query and discovers and assembles the image set itself.Visual storytelling work includes Huang et al. (2016) and Hu et al. (2020).Source: §6.2 · p. 9
  • IBC fills the gap in existing retrieval and generation tasks by making bundle-level consistency the criterion for discovering a relationally coherent image set from an unstructured pool.Its validity depends on interactions among images and joint relational, temporal, and narrative constraints, not only individual query relevance.Source: §6.1 · pp. 8–9, §6.2 · p. 9

7 Conclusion

The paper introduces Image Bundle Composition (IBC), a paradigm for composing relationally cohesive image bundles, and presents IBCBench and BundleWeaver to establish and address this task.

7 Conclusion

  • BundleWeaver significantly outperforms existing baseline paradigms by treating IBC as query-conditioned incremental hyperedge discovery.It combines adaptive subquery generation, parallel beam search, and whole-bundle VLM reranking to enforce cross-image structural constraints.Source: §7 · p. 9
  • IBCBench is introduced as the first meticulously verified benchmark dataset for Image Bundle Composition.The benchmark is intended to establish the new task while mitigating the computational bottlenecks of combinatorial annotation.Source: §7 · p. 9
  • Image Bundle Composition shifts retrieval from ranking isolated snapshots to dynamically composing cohesive bundles bound by explicit cross-image relations.Source: §7 · p. 9

Limitations … C.1 Property Requirements of a Valid Image Bundle

The paper presents BundleWeaver as a foundational starting point for a difficult retrieval task, while identifying current scope limitations and defining strict criteria for valid image bundles.

Limitations

  • The study covers static images in YFCC personal photo collections, not specialized domains, continuous video, or end-to-end fine-tuning for IBC.The authors identify parameter-efficient tuning tailored to IBC as a future direction.Source: Limitations · p. 9

A Ethical Considerations

  • IBCBench uses YFCC-100M images under Creative Commons licenses authorizing research, audits spatiotemporal metadata, and will follow original redistribution guidelines.Source: A Ethical Considerations · p. 12
  • The authors say privacy risks are minimized by public licensed assets and relational reasoning rather than identity profiling, envisioning personal-gallery use rather than unauthorized surveillance.They describe use in private galleries or encrypted local storage.Source: A Ethical Considerations · p. 12

B Impact Discussion

  • BundleWeaver improves Exact Match over the strongest baseline, but its absolute performance remains modest, leaving substantial headroom for this combinatorial task.The authors position it as a foundational baseline rather than a definitive endpoint.Source: B Impact Discussion · p. 12

C.1 Property Requirements of a Valid Image Bundle

  • A valid bundle requires every image to be indispensable and the query to impose a higher-order relation across images, not merely list independent targets.Source: §C.1 · p. 12
  • A valid bundle must be unambiguous against the full pool, with no smaller subset satisfying the query equally well or better, and must avoid near-duplicate images.Source: §C.1 · p. 12

C.2 Source Data and Visual Profiles … D.1 Prompt Demonstration

This section describes the construction and curation of IBCBench, including relational bundle verification, human review, dataset composition, and prompts for visual profiling and validation. It also specifies implementation details for checking query generation and for BundleWeaver’s incremental search and coherent-set scoring and comparison.

C.2 Source Data and Visual Profiles

  • The pool draws on 57 YFCC100M users’ 109,467 photos with valid EXIF timestamps and over 98.3% GPS coverage, excluding user-defined album IDs.GPT-4o generated captions, word-level tags, and event signals for all images; RzenEmbed multimodal embeddings helped penalize redundancy during candidate mining.Source: §C.2 · p. 12

C.3 Spatiotemporal Candidate Mining

  • Sessions split at gaps over 6 hours or jumps over 20 km; sliding windows of 3–5 images scored below 2.5 were discarded, leaving 7,460 candidates.The composite heuristic considers spatiotemporal coverage, semantic richness, and redundancy; coverage rewards duration, distance, and shared reverse-geocoded address prefixes, while semantic richness favors diverse tags and event signals.Source: §C.3 · p. 13

C.4 VLM-Driven Bundle Verification and Query Generation

  • A VLM accepted candidates only when they fit a cross-image template, yielded a natural-language query, and named a specific shared anchor; three programmatic gates rejected forbidden styles and vague anchors.The VLM considered same-location dynamics and cross-location structures; only about 8.5% of candidates passed verification.Source: §C.4 · pp. 13–14

C.5 Human Review

  • Of 1,674 candidates reviewed, 667 (39.8%) were accepted; accepted cases were cross-validated and disputed cases discussed, while inspected false negatives were negligible.Four expert annotators checked constraints C1–C4, refined query naturalness, and rejected queries lacking pool-wide uniqueness (C3).Source: §C.5 · pp. 14–15

C.6 Final Dataset Statistics

  • IBCBench contains 667 queries over 109,467 images, with bundle sizes of 3 (24.3%), 4 (32.2%), and 5 (43.5%) images and relations split between same-location (52.5%) and cross-location (47.5%).The pipeline’s end-to-end acceptance rate was under 9%.Source: §C.6 · p. 15

C.7.1 Prompt for Building Visual Profiles

  • The visual-profiling prompt requests English JSON containing a detailed concise caption, short tags, possible event or state-change clues, and a note on bundle utility.The note asks whether an image may help a time, state, route, or event progression bundle.Source: §C.7.1 · p. 15

C.7.2 Prompt for Bundle Verification and Query Generation

  • The validation prompt requires one allowed template, a concrete instance-level anchor, jointly necessary images, and a unique relational query rather than independent image targets.Its templates distinguish same-location change or episode progression from cross-location repeated rituals, arrival moments, routes, activities, or specific shared patterns.Source: §C.7.2 · p. 15
  • The query-generation prompt records pass/fail judgments for single-image insufficiency, non-substitutability, low redundancy, natural-language query validity, template fit, and review risk.Template fit requires a brief explanation of why the template ID applies; review risk is labeled low, medium, or high.Source: §C.7.2 · p. 16
  • When rejecting an output, the prompt requires accept=false and query=null, plus the closest template, any identifiable shared anchor, and a specific rejection reason.Source: §C.7.2 · pp. 16–17

C.7.3 Details of Programmatic Gates

  • Programmatic gates recheck accepted outputs for forbidden query patterns, sufficiently specific shared anchors, and template-family membership, overriding model acceptance on any violation.Anchors shorter than four tokens or dominated by generic phrases from a 28-term blocklist are rejected.Source: §C.7.3 · p. 17

D Implementation Details

  • BundleWeaver uses RzenEmbed with FAISS IndexFlatIP, five seeds, beam width four, five candidates per step, 24-hour and 50-km pruning bounds, λ=0.3, and GPT-4o for generation and reranking.Experiments use the full IBCBench dataset of 667 queries and bundles and run on two A100 GPUs.Source: D Implementation Details · p. 17

D.1.1 Prompt for Adaptive Sub-Query Generation

  • Adaptive sub-query generation receives the original query, prior search directions, and found-image captions, then outputs one SEARCH instruction for a missing aspect or DONE when the bundle is complete.The prompt requires the output to be exactly SEARCH: <concise description of the photo to find next> or DONE, with nothing else.Source: §D.1.1 · p. 17

D.2 Baseline Implementation … E.6 Case Study

The additional experiments characterize BundleWeaver as a slower, iterative alternative to atomic retrieval and examine reranking, cardinality, hyperparameters, human validation, and case-study behavior. The case studies show how bundle-level reasoning can preserve spatial and temporal relations that isolated retrieval or static decomposition misses.

D.2 Baseline Implementation

  • The baselines span independent multimodal or caption retrieval, metadata heuristics, and VLM decomposition into subqueries followed by reranking.The strongest baseline uses RzenEmbed to retrieve up to five images per subquery, then asks a VLM to select one per group; metadata variants exploit user, session, time, or location.Source: §D.2 · p. 18

E.1 Efficiency Analysis

  • RzenEmbed took 0.3s per query, Session Clustering 0.8s, and GPT-4o decompose-and-rerank 12s; BundleWeaver adds latency for iterative reasoning and verification.The paper frames this extra cost as a trade-off for addressing IBC, which atomic retrieval methods are insufficient to solve.Source: §E.1 · p. 19, Table 4 · p. 19

E.2 Reranking Strategy Comparison

  • Pointwise VLM scoring yielded the best F1 and EM, while listwise selection achieved competitive Recall; the authors attribute this to less attention dilution and better scalability.Pointwise evaluation audits bundles independently, whereas listwise prompts may present dozens of images simultaneously.Source: §E.2 · p. 19, Table 5 · p. 19

E.3 Expanded Baseline Evaluation

  • With dynamic cardinality, strict similarity thresholds missed bundle members and loose thresholds admitted distractors, so oracle-size evaluation favors atomic retrieval baselines.The authors conclude the BundleWeaver gap should be larger without oracle cardinality information.Source: §E.3 · pp. 19–20, Table 6 · p. 19

E.4 Hyperparameter Study

  • BundleWeaver performed best at beam width Bu = 4 and C = 5; temporal windows of 12–24h and radii of 20–100km maintained performance, with defaults T = 24h and R = 50km.Over-expansion introduced noisy candidates, while windows that were too tight excluded valid longer events or cross-location bundles.Source: §E.4 · p. 20, Figure 6 · p. 20, Figure 7 · p. 20

E.5 Human Validation of Model Results

  • Only 2% of BundleWeaver predictions were judged reasonable alternative bundles and none of Claude-Sonnet-4.5’s were; BundleWeaver also had higher Precision, Recall, and F1 in shared failure cases.The audit of 100 queries suggests missing valid alternatives did not substantially distort strict set evaluation.Source: §E.5 · p. 20, Table 7 · p. 20

E.6 Case Study

  • Across route and wedding case studies, BundleWeaver maintained spatial continuity and single-wedding identity, unlike atomic retrieval and static decomposition.Embedding retrieval returned visually redundant images, while decomposition retrieved landmarks or event components from unrelated trips or weddings.Source: §E.6 · p. 21

E.7 Failure Case Analysis · F Discussions on Potential Concerns

BundleWeaver can retrieve relevant images yet fail to compose complete, consistent, coherently ordered bundles. The analysis identifies four failure modes and points to stronger identity-aware representations, temporal modeling, and adaptive verification as future directions.

E.7 Failure Case Analysis

  • Rare, weakly observable, or semantically ambiguous roles can remain uncovered when missing-role queries are insufficient or the candidate pool lacks distinguishable examples.Visually relevant images may still fail to cover every role required by the underlying event.Source: §E.7 · p. 21
  • Locally relevant images from different users or occasions can be combined into an inconsistent narrative when identity cues are insufficient.Source: §E.7 · pp. 21–22
  • Semantically related images may be ordered incorrectly, with skipped stages, distractors, or reversed transitions because temporal relations are weakly represented in image embeddings.Source: §E.7 · p. 22
  • Images close in time or location may belong to unrelated activities, so metadata can narrow search but cannot independently establish relational relevance.Source: §E.7 · p. 22
  • IBC retrieval requires comprehensive role coverage, global consistency, and coherent event-level reasoning beyond visual similarity; future work may strengthen identity, temporal modeling, and verification.The suggested improvements target identity-aware representations, more reliable temporal modeling, and adaptive verification for ambiguous bundles.Source: §E.7 · p. 22

F Discussions on Potential Concerns

F.1 Is contextual candidate pruning a benchmark bias? … G.2 Lemma 1: Strict Non-Submodularity of Relational Bundles

The paper argues that contextual pruning reflects query-specified physical constraints rather than a shortcut, while its evaluation and theoretical analysis support the need for relational bundle reasoning. It reports low rates of alternative valid answers and proves that IBC’s relational objective is strictly non-submodular.

F.1 Is contextual candidate pruning a benchmark bias?

  • CCP uses timestamps and GPS as query-faithful physical constraints, but defines only a broad feasible region and does not determine the answer.Queries may specify a trip, same-day progression, route, or nearby repeated activity; ignoring such context would require unrealistic global search.Source: §F.1 · p. 22
  • CCP lowers RzenEmbed F1 from 14.88 to 8.80 and GPT-4o Decompose-and-Rerank F1 from 17.84 to 16.74; fewer than 9% of mined windows survive verification.Locality can recover complete bundles but also introduce relationally irrelevant distractors that hurt partial set recovery.Source: §F.1 · p. 22
  • 24.69 F1 without CCP substantially exceeds the strongest GPT-4o static baseline’s 17.84 F1, showing BundleWeaver remains strong without pruning.The paper attributes its core gains to adaptive missing-role reasoning, beam-based bundle construction, and whole-bundle verification.Source: §F.1 · pp. 22–23

F.2 Is the "uniqueness" assumption fully validated?

  • False-negative rates were 2% for BundleWeaver and 0% for Claude; human verification screened alternatives and rejected queries with interchangeable image subsets.Reviewers used a nearest-neighbor visual panel to search for alternatives, while structural constraints could refine a query into a unique trajectory.Source: §F.2 · p. 23

G Theoretical Limitations of Atomic Retrieval in IBC

  • The theoretical analysis argues that static Decompose-and-Rerank suffers relational blindness because IBC’s non-submodularity prevents local top-k retrieval from converging to the global optimum.The section uses the non-submodularity result to establish the stated limitation of local top-k retrieval.Source: G Theoretical Limitations of Atomic… · p. 23

G.1 Preliminaries and Problem Definition

  • IBC scoring combines independent image-to-subquery semantic matches with a non-decomposable bundle-level relational constraint, treated as a hard penalty when violated.The relational indicator covers constraints such as spatiotemporal continuity and identity consistency.Source: §G.1 · p. 23

G.2 Lemma 1: Strict Non-Submodularity of Relational Bundles

  • A bundle image yields greater marginal gain when added to a relationally complete story than to an incomplete one, violating diminishing returns and strict submodularity.The corollary states that greedy selection based on independent marginal gains cannot guarantee a constant-factor approximation, motivating adaptive context-aware expansion.Source: §G.2 · p. 24

G.3 Theorem 1: The Relational Blindness Bound of Independent Top-k

Theorem 1 shows that independent top-k retrieval can make the optimal bundle impossible to discover when locally stronger distractors violate global relations. Expanding k beyond the distractor count avoids exclusion only at the cost of an exponentially growing search space.

G.3 Theorem 1: The Relational Blindness Bound of Independent Top-k

  • When k ≤ M, the probability of discovering the optimal bundle through Decompose-and-Rerank is strictly 0 under the theorem’s dominant-distractor assumptions.For each target image, M distractors score higher on the local subquery, and any bundle containing a distractor violates the global relational constraint.Source: §G.3 · p. 24
  • Because atomic retrieval cannot see global constraints, top-k selection excludes each target image, leaving the VLM to search only bundles containing dominant distractors.The VLM is restricted to the Cartesian product of the independently retrieved candidate sets.Source: §G.3 · p. 24
  • Avoiding this failure requires k > M, but the resulting VLM search space O(k^K) grows exponentially, making the independent-splitting strategy computationally intractable.The theorem characterizes this failure as Relational Blindness.Source: §G.3 · p. 25
  • The theorem explains the reported catastrophic Exact Match drop: wedding photos from different events act as high-similarity distractors that displace identity-consistent images needed to satisfy global relations.The passage points to Case Study 2, Figure 8, for examples of these results.Source: §G.3 · p. 25, Figure 8 · p. 26

Related papers

Loading 2608.28695v1…