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 retrieval often scores images independently, but IBC targets cohesive bundles whose members jointly satisfy relational intent. The paper introduces IBCBench and BundleWeaver, which incrementally discovers query-conditioned hyperedges and verifies complete bundles. BundleWeaver achieves state-of-the-art performance across all metrics, while removing whole-bundle VLM reranking causes a performance drop.

  • Problem

    Atomic retrieval fails to capture visual stories whose images are bound by higher-order relations, and IBC therefore requires non-decomposable joint relevance.

  • Method

    The paper constructs IBCBench and proposes BundleWeaver, which incrementally discovers query-conditioned hyperedges using adaptive search and whole-bundle VLM reranking.

  • Results

    BundleWeaver achieves state-of-the-art performance across all metrics, while removing VLM pointwise reranking causes a performance drop.

  • Takeaways & Limitations

    The results support explicitly modeling cross-image composition rather than relying only on independent image-level matching.

  • Takeaways & Limitations

    IBC faces combinatorial search and non-decomposable joint relevance, making exhaustive subset selection intractable and independent top-K selection insufficient.

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 paper introduces Image Bundle Composition (IBC) to retrieve cohesive visual stories rather than independently ranked images. It also presents IBCBench and BundleWeaver to support benchmarked, relational bundle discovery in unstructured photo collections.

  • Traditional text-to-image retrieval scores candidate images independently, producing ranked lists of individually relevant items.
  • Human search in personal photo collections often targets cohesive visual stories whose heterogeneous images are connected by temporal, event-based, or spatial relations.
  • IBC shifts retrieval from ranking isolated images to dynamically composing cohesive bundles from a vast, unstructured photo pool.
  • IBC requires navigating a combinatorially explosive search space because target bundles are not predefined or indexed.
  • IBCBench is the first IBC benchmark, containing 109,467 images and 667 verified queries constructed through a semi-automated pipeline.
  • BundleWeaver treats IBC as query-conditioned hyperedge discovery, using adaptive role search, incremental expansion, and whole-bundle verification.

2 Task Formulation

The paper formalizes IBC as selecting a compact subset whose joint relevance captures relations across images, unlike traditional point-wise retrieval. It identifies combinatorial explosion and non-decomposable relevance as central challenges.

  • 2.1 Preliminaries: Traditional Text-to-Image Retrieval: Traditional text-to-image retrieval independently scores each candidate image against a query and treats relevance as decoupled across the retrieved list.
  • 2.2 Image Bundle Composition (IBC): IBC seeks a compact image bundle that collectively fulfills the query’s relational or narrative intent rather than a single image or disconnected list.
  • 2.2 Image Bundle Composition (IBC): The candidate bundle is a subset of distinct images with dynamically determined size bounded by Kmax.
  • 2.2 Image Bundle Composition (IBC): IBC optimizes a joint relevance function Φ(B, q) that evaluates structural, temporal, or spatial relations across the bundle as a whole.
  • 2.3 Discussions on Challenges of IBC: IBC introduces two fundamental challenges: combinatorial explosion and non-decomposability of joint relevance.
  • 2.3 Discussions on Challenges of IBC: The bundle search space grows combinatorially with pool size N, making exhaustive evaluation computationally intractable for typical personal photo collections.
  • 2.3 Discussions on Challenges of IBC: Joint relevance cannot be approximated by aggregating independent image-level scores, so selecting the top-K images can fail to satisfy the intended relation.
  • 2.3 Discussions on Challenges of IBC: The paper provides additional theoretical analysis of atomic retrieval’s limitations for IBC in Appendix G.

3 Dataset Construction

IBCBench addresses the difficulty of annotating relational image bundles through a semi-automated construction process with strict acceptance criteria. The dataset is designed for cohesive, higher-order bundle retrieval.

  • IBCBench is introduced because exhaustive annotation is computationally impossible and unconstrained VLM generation can produce ambiguous or decomposable image lists.
  • The dataset construction uses a five-stage semi-automated pipeline to identify and verify candidate bundles.
  • A valid candidate bundle must satisfy strict criteria including joint completeness and cross-image binding.

A. Task Reframing

The task is framed around producing a cohesive image bundle.

  • The stated goal is a cohesive image bundle.
  • Cohesion is the defining target named for the image bundle.
  • The task’s goal centers on the bundle as a unified visual output.

B. Seed Initialization

The benchmark construction pipeline mines diverse candidate windows from a large photo pool, verifies them against relational templates, and subjects them to expert review. The resulting IBCBench contains 667 verified queries with varied bundle sizes and balanced relation types.

  • Candidate Mining: 109,467 YFCC-100M photos with valid spatiotemporal metadata were grouped into sessions and pruned into 7,460 diverse candidate windows.Dense captions, tags, multimodal embeddings, and sliding windows of size K ∈[3, 5] supported candidate mining.
  • VLM Verification & Human Review: Claude-Opus-4.5 verified candidates against same-location and cross-location relational templates before four expert annotators reviewed them against C1-C4.The review filtered ambiguous cases and checked global uniqueness.
  • Final Dataset Statistics: 667 high-quality queries remained after an end-to-end acceptance rate below 9%.The final benchmark contains bundles of 3, 4, or 5 images.
  • Final Dataset Statistics: 52.5% of queries focus on same-location dynamics, while 47.5% capture cross-location structural ties.These proportions indicate a balanced distribution across the two reported semantic relation categories.

4 Methodology

BundleWeaver treats image-bundle retrieval as query-conditioned hyperedge discovery and searches incrementally rather than enumerating subsets. Its pipeline combines diverse seeds, adaptive beam-search expansion, and whole-bundle VLM reranking.

  • Problem Formulation: BundleWeaver reformulates IBC as query-conditioned hyperedge discovery over an implicit hypergraph whose hyperedges are coherent image bundles.Because valid hyperedges are not predefined, the system constructs them dynamically from the image pool.
  • Diverse Seed Initialization: The LLM extracts a visual anchor and initial search direction, which a dense multimodal model uses for global nearest-neighbor retrieval.This process anchors initialization to promising vertices in the unstructured image pool.
  • Diverse Seed Initialization: Diverse Seed Initialization selects Ns semantically and visually diverse seeds from the top-K retrieved images, creating independent roots for parallel construction.The strategy reduces near-duplicate starting points and local-search collapse.
  • Adaptive Hyperedge Expansion via Beam Search: At each beam-search step, contextual pruning restricts candidates to plausible spatiotemporal neighborhoods, while the LLM generates sub-queries for missing relational elements.The expansion searches for complementary images rather than simply adding locally relevant ones.
  • Adaptive Hyperedge Expansion via Beam Search: Path scoring balances step-wise local matching against holistic bundle alignment through λ, which controls their trade-off.The method retrieves top-C local candidates and uses the composite score to decide which branches survive.
  • Whole-Bundle VLM Reranking: A VLM reranks completed paths by evaluating all images with the query for structural coherence and cross-image binding, returning the highest-scoring hyperedge.Beam search terminates when the LLM judges a bundle complete, subject to maximum depth Kmax.

5 Experiments

The experiments evaluate IBC with set-level metrics and compare atomic, metadata, static decomposition, and incremental composition paradigms. BundleWeaver performs best across reported analyses, while ablations show that diverse seeding, contextual pruning, parallel beam search, and whole-bundle verification support relationally coherent retrieval.

  • Experimental Setup: IBC is evaluated using set-level Precision, Recall, F1, and Exact Match, averaged across queries.Exact Match measures the percentage of perfectly predicted ground-truth bundles.
  • Experimental Setup: The benchmark compares multimodal embedding, caption-plus-text embedding, heuristic metadata augmentation, and VLM decompose-and-rerank baselines.These baselines represent atomic retrieval, text-mediated retrieval, metadata heuristics, and static agentic decomposition.
  • Experimental Setup: Oracle-size evaluation truncates ranked lists to the ground-truth bundle size, making predicted and ground-truth set Precision, Recall, and F1 mathematically equivalent.This protocol applies to baselines that naturally return ranked lists of individual images.
  • Main Results: Traditional point-wise matching loses cross-image relations, while naive metadata augmentation degrades F1 because spatiotemporal proximity alone cannot satisfy query-specific relational roles.The results argue against solving IBC through independent text-image alignment or rigid locality heuristics.
  • Main Results: Static decomposition improves coverage but retains very low Exact Match because independently retrieved subqueries cannot enforce inter-image constraints.BundleWeaver instead achieves state-of-the-art performance across all metrics through adaptive hyperedge expansion and whole-bundle reranking.
  • Component Ablation: Ablations show that removing diverse seeding, contextual candidate pruning, parallel beam search, or VLM pointwise reranking reduces performance.The reported explanations link these components to search coverage, candidate tractability, resistance to early errors, and bundle-level identity or structural verification.
  • Breakdown Analysis: BundleWeaver remains stronger as bundle size increases and across location types, with a smaller performance drop than baselines and better robustness on cross-location bundles.RzenEmbed performs notably worse on cross-location bundles, whereas BundleWeaver achieves the best metrics for both location types.
  • Backbone Generalizability: Across frontier VLM backbones, BundleWeaver consistently outperforms decompose-and-rerank and produces a large Exact Match increase.The comparison attributes the gain to incremental hyperedge discovery coupled with whole-bundle verification rather than to a particular VLM backbone.

6 Related Works

Related retrieval tasks have added multiple images, personal photo streams, visual references, or iterative reasoning, but they do not explicitly retrieve relationally coherent bundles from unstructured image pools. IBC defines bundle-level consistency as the retrieval criterion and requires joint reasoning over cross-image relations.

  • Existing Retrieval Paradigms: Conventional vision-language retrieval independently estimates relevance between a query and each candidate image.MSCOCO and Flickr30K established this point-wise text-image and image-text retrieval paradigm.
  • Existing Retrieval Paradigms: Retrieval models have progressed from dual encoders such as CLIP and SigLIP to MLLM-based representations and agentic methods with iterative reasoning, decomposition, and verification.These developments expand model expressiveness and retrieval procedures without changing the basic comparison here.
  • Image Bundle Composition: IBC shifts the target from isolated image relevance to cohesive bundles whose images jointly satisfy relational, temporal, and narrative constraints.Its validity depends on interactions among images rather than independent cross-modal matching.
  • Related Tasks: Multi-image retrieval, lifelog retrieval, and composed image retrieval address related collection or multi-image settings but differ from IBC in their stated objectives.These tasks do not explicitly establish relationally coherent bundle retrieval from a large unstructured image pool in the supplied comparison.
  • Related Tasks: Visual storytelling assumes the images are already given and generates textual narratives, whereas IBC starts from text and discovers the image set.The distinction is therefore between narrating a supplied collection and retrieving a collection that fulfills the query.

7 Conclusion

The paper introduces IBC, IBCBench, and BundleWeaver to support retrieval of cohesive image bundles under cross-image structural constraints. BundleWeaver uses incremental hyperedge discovery with adaptive subqueries, parallel beam search, and whole-bundle VLM reranking, outperforming existing baseline paradigms.

  • Conclusion: IBC reframes retrieval as dynamically composing cohesive bundles bound by explicit cross-image relations.The paradigm moves beyond ranking isolated snapshots.
  • Conclusion: IBCBench is introduced as a verified benchmark for evaluating this new retrieval paradigm.The dataset addresses computational bottlenecks associated with combinatorial annotation through dedicated benchmark construction.
  • Conclusion: BundleWeaver reformulates IBC as query-conditioned incremental hyperedge discovery using adaptive subquery generation, parallel beam search, and whole-bundle VLM reranking.These components are presented as enforcing cross-image structural constraints.
  • Conclusion: BundleWeaver significantly outperforms existing baseline paradigms on IBC.The conclusion attributes this result to its bundle-level treatment of cross-image constraints.

Limitations

The paper identifies scope, evaluation, and efficiency limitations that bound current IBC and BundleWeaver claims. It also reports that the benchmark remains difficult despite verification and that several design choices involve practical trade-offs.

  • Scope: The study is limited to static images from personal photo collections and does not yet cover specialized domains or dynamic modalities such as video.The authors specifically mention medical imaging sequences, legal evidentiary archives, and Video Bundle Retrieval as unexplored settings.
  • Methodological boundary: BundleWeaver operates training-free and zero-shot with off-the-shelf foundation models, leaving end-to-end fine-tuning unexplored.This demonstrates generalized relational reasoning but does not establish the effect of task-specific training.
  • Evaluation: Absolute Exact Match performance remains modest despite relative improvement over the strongest baseline, leaving substantial headroom for exact bundle recovery.The paper attributes this difficulty to recovering non-decomposable visual narratives from a massive O(NK) combinatorial space.
  • Efficiency: BundleWeaver incurs additional latency because iterative agentic reasoning and verification replace one-shot embedding matching.This is an explicit trade-off between richer exploration and query efficiency.
  • Design sensitivity: Performance depends on search and mining choices, including temporal and spatial windows, beam width, candidate count, and pruning thresholds.Tight windows can exclude valid bundles, excessive expansion can degrade retrieval, and heuristic pruning reduces the candidate pool to 7,460 windows.

E.7 Failure Case Analysis

BundleWeaver can fail despite modeling relational completeness and bundle-level consistency. The analysis identifies failures involving missing roles, inconsistent identities or events, broken temporal order, and misleading metadata neighborhoods.

  • Incomplete relational coverage: BundleWeaver may retrieve visually relevant images while failing to cover rare, weakly observable, or unavailable event roles.Generated missing-role queries remain insufficient when the required role is absent from the candidate pool or difficult to recognize.
  • Identity and event inconsistency: Images satisfying local subqueries can still violate global identity or event consistency when different occasions share similar scenes or objects.Without sufficient identity cues, such images may be combined into an inconsistent narrative.
  • Broken temporal or narrative order: Semantically related images may be ordered incorrectly, skipping stages, adding distractors, or reversing intended event transitions.Temporal relationships are often weakly represented in individual image embeddings.
  • Metadata-neighborhood distractors: Metadata proximity can introduce misleading candidates because nearby images may belong to unrelated activities or different events.Metadata is useful as an auxiliary signal but cannot independently resolve IBC retrieval.
  • Overall implication: Successful IBC retrieval requires comprehensive role coverage, global consistency, and coherent event-level reasoning beyond visual similarity.The paper points to identity-aware representations, stronger temporal modeling, and adaptive verification as future directions.

F.1 Is contextual candidate pruning a benchmark bias?

The paper argues that contextual candidate pruning is a query-faithful physical constraint rather than a benchmark shortcut, while showing that locality alone is insufficient. It also supports exact set matching through uniqueness checks and theoretically explains why static retrieval is relationally blind.

  • CCP as a query-faithful constraint: CCP reflects query-implied temporal and geographic constraints, such as trips, same-day progressions, routes, or nearby-venue activities.The paper compares this use of metadata to date or location filtering in personal photo search.
  • CCP is insufficient alone: Adding CCP decreases RzenEmbed F1 from 14.88 to 8.80 and GPT-4o Decompose-and-Rerank F1 from 17.84 to 16.74, although EM improves slightly.The result indicates that physical locality can introduce relationally irrelevant distractors.
  • Dataset construction evidence: Fewer than 9% of spatiotemporally mined candidate windows survive semantic verification and human review.Thus, most physically plausible windows are not valid IBC answers.
  • BundleWeaver without CCP: The w/o-CCP BundleWeaver variant achieves 24.69 F1 versus 17.84 F1 for the strongest GPT-4o static baseline.The paper attributes the core gain to adaptive missing-role reasoning, beam-based construction, and whole-bundle verification.
  • Uniqueness and evaluation: Manual inspection found a 2% false-negative rate among 100 sampled automatic-evaluation errors.A reasonable alternative had to satisfy all stated constraints and the query while using different images.
  • Uniqueness enforcement: Queries with interchangeable subset combinations are rejected, while structural trajectories such as a ceremony progression are accepted as unique.Nearest-neighbor review helps annotators search for alternative bundles during verification.
  • Theoretical account: IBC’s joint relevance function is strictly non-submodular under relational structural constraints.A complete narrative can create a larger marginal gain for an image after additional context is present, violating diminishing returns.
  • Theoretical implication: Because local marginal-gain selection lacks a constant-factor guarantee, IBC requires adaptive, context-aware expansion rather than independent greedy retrieval.This theoretical result motivates Bundle-Search-style strategies.

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

Theorem 1 formalizes when independent top-k retrieval cannot discover the optimal IBC bundle: dominant distractors locally outperform target images while violating global relations. With cutoff k ≤ M, the optimal bundle’s discovery probability is zero, and expanding k causes exponential search growth.

  • Assumptions: Theorem 1 assumes each target image has M dominant distractors with higher local alignment but invalid global relational membership.These distractors are defined relative to each target image and its subquery.
  • Relational blindness bound: When k ≤ M, independent top-k retrieval excludes every target image from its local candidate space, making optimal-bundle discovery probability strictly 0.The atomic scorer cannot observe the global relational constraint during independent retrieval.
  • Restricted search space: The subsequent VLM search is restricted to the Cartesian product of local candidate spaces, so every examined bundle contains at least one dominant distractor.The candidate space is Sglobal = S1 × S2 × · · · × SK.
  • Consequence: The resulting composite has joint relevance significantly below the optimum because its candidates violate the global relational constraint.The theorem describes this as a fragmented composite forced by independent retrieval.
  • Computational trade-off: Avoiding the bound by choosing k > M makes the VLM search space O(k^K), which grows exponentially and becomes computationally intractable.Thus, simply increasing local retrieval cutoffs does not provide a practical solution.
  • Empirical interpretation: The analysis connects this mechanism to visually similar wedding photos from different events displacing identity-consistent images in local rankings.These cross-event images act as dominant distractors in the reported case study.
Loading 2608.28695v1…