Weaving Visual Narratives: Agentic Image Bundle Composition Beyond Atomic Visual Matching¶
Conference: EMNLP 2026
Paper: ACL Anthology
Area: Multimodal VLM
Keywords: Image Bundle Composition, Agentic Retrieval, Hyperedge Discovery, Multimodal Benchmark, Visual Storytelling
TL;DR¶
Addressing the fundamental blind spot of conventional point-wise image retrieval that evaluates candidate images in isolation, this paper introduces the Image Bundle Composition (IBC) paradigm, constructs the meticulously audited IBCBench dataset, and develops BundleWeaver—an agentic framework modeling bundle composition as incremental hyperedge discovery with whole-bundle VLM verification, yielding a 380% relative gain in Exact Match over frontier baselines.
Background & Motivation¶
Image retrieval has conventionally operated under a strict point-wise matching paradigm, where given a natural language query, each candidate photograph in an unstructured pool is independently scored to produce a ranked list of isolated visual assets. While contemporary vision-language models and dense multimodal embedding architectures have significantly pushed the envelope of single-image semantic alignment and even multi-hop historical navigation, their evaluation and operational primitives remain fundamentally decoupled. In everyday life, however, human memory and search intent within personal photo archives rarely revolve around isolated snapshots; instead, individuals recall and seek compact visual stories bound by intricate structural ties. A user frequently looks for a cohesive journey documenting a mountain trek across key landmarks, the dynamic temporal progression of an entire wedding ceremony, or the subtle shifting atmosphere of a cityscape from dawn to dusk. In these realistic scenarios, individual images may exhibit marked visual heterogeneity and are rendered truly meaningful only as an integrated collective whole.
This practical disparity exposes a severe structural tension between atomic retrieval algorithms and genuine narrative search intent: the joint relevance function of an image bundle is strictly non-decomposable, meaning that the overall narrative score of an image subset cannot be formulated as a monotonic aggregation of individual image-query similarities. Because target image bundles are neither indexed a priori nor pre-clustered, discovering an optimal combination across an uncurated personal gallery of size \(N\) poses a prohibitive combinatorial search space scaling as \(O(N^{K_{\max}})\). Standard divide-and-conquer agentic approaches that decompose the prompt into multiple sub-queries inevitably fall victim to "relational blindness." When sub-queries are executed independently across the global pool, atomic scorers are easily fooled by "dominant distractors"—images that achieve exceptionally high individual similarity to a sub-query but originate from completely different events, users, or physical settings. As a result, static decompose-and-rerank pipelines assemble fragmented, disjointed "Frankenstein" collections that violate the fundamental narrative continuity demanded by the user.
Establishing a viable solution to this bottleneck requires overcoming two intertwined challenges: formulating an unambiguous benchmark dataset where candidate bundles represent provably unique, non-decomposable answers within a massive pool, and designing an efficient navigation framework capable of actively pursuing missing narrative roles without exhaustive subset enumeration. Core idea: reformulate Image Bundle Composition as query-conditioned incremental hyperedge discovery on an implicit multimodal graph, dynamically deducing missing relational roles via an LLM agent during parallel beam expansion and enforcing global narrative integrity via whole-bundle VLM reranking.
Method¶
Overall Architecture¶
BundleWeaver conceptualizes the unindexed image pool as the vertex set of an implicit hypergraph, wherein a valid, cohesive image bundle constitutes an unindexed hyperedge connecting a discrete subset of vertices through higher-order semantic and spatiotemporal relations. To bypass the intractable \(O(N^K)\) combinatorial explosion, the framework models hyperedge discovery as a sequential, query-conditioned pathfinding process across three coordinated stages. First, Diverse Seed Initialization identifies multiple semantically orthogonal and topographically dispersed visual anchors to serve as parallel root vertices across the global embedding space. Second, starting from these diverse seeds, Adaptive Hyperedge Expansion executes a parallel beam search over a contextually pruned candidate pool; at each iterative step, an LLM reasoning agent examines the partial narrative trajectory and deduces an adaptive sub-query targeting the exact missing narrative role, while a dense multimodal scoring function balances step-wise alignment with overall trajectory momentum. Finally, upon reaching natural narrative termination, all candidate hyperedges are gathered into a unified pool for Whole-Bundle VLM Reranking, where a vision-language model evaluates the complete image sets holistically to establish the optimal, fully bound visual bundle \(B^*\).
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Natural Language Query + 109K Image Pool"] --> B["Diverse Seed Initialization<br/>LLM Anchor Extraction & Multi-Peak Seed Selection"]
B --> C["Contextual Candidate Pruning<br/>Spatiotemporal Feasibility Gating from Seeds"]
C --> D["Adaptive Hyperedge Expansion<br/>LLM Missing-Role Deduction & Parallel Beam Search"]
D -->|Termination or Max Depth| E["Whole-Bundle VLM Reranking<br/>Holistic Coherence & Identity Verification"]
E --> F["Optimal Cohesive Image Bundle B*"]
Key Designs¶
1. Diverse Seed Initialization: Multi-Peak Grounding to Circumvent Local Traps
Directly initiating an iterative graph expansion from a single top-1 nearest-neighbor retrieval result severely restricts the search scope, causing the agent to collapse into local optima or retrieve redundant starting points. To establish robust global coverage, BundleWeaver prompts an LLM to dissect the original query \(q\), extract the central visual anchor, and formulate an initial dense search vector \(d_1\). After performing an initial global approximate nearest neighbor search over the \(N\)-image corpus using dense embeddings, the framework selects a subset of \(N_s = 5\) seed images optimized for maximum mutual semantic and visual diversity. Each seed image serves as the independent root for a parallel hyperedge construction tree, ensuring that distinct candidate narrative starting points are explored concurrently without premature commitment.
2. Contextual Candidate Pruning: Physical Feasibility Boundaries Against Combinatorial Drift
Real-world human experiences inherently follow coarse physical constraints governed by temporal progression and geographic locality. Exposing the reasoning agent to the entire unconstrained pool at every intermediate step causes the search tree to be inundated with visually similar yet physically impossible distractors. To maintain computational tractability while remaining faithful to user intent, BundleWeaver introduces Contextual Candidate Pruning (CCP). For any search branch rooted at seed image \(x_1\), the permissible expansion pool for subsequent bundle members is dynamically bounded to a spatiotemporal neighborhood within time window \(\tau_t \le 24\) hours and geographic radius \(\tau_g \le 50\) km. Crucially, CCP acts strictly as a coarse plausibility filter rather than a solution; images captured in close proximity frequently represent completely unrelated activities, preserving the necessity of subsequent relational reasoning.
3. Adaptive Hyperedge Expansion: Iterative Missing-Role Deduction and Balanced Beam Scoring
Because the joint relevance function of an image bundle is strictly non-submodular, standard greedy subset selection based on isolated marginal gains is mathematically incapable of guaranteeing a bounded approximation. BundleWeaver resolves this by implementing a parallel beam search of width \(B_u = 4\). At expansion step \(k\), given the partially constructed bundle \(B_{k-1}\), standard embedding models cannot deduce what structural component remains unsatisfied. An LLM reasoning agent is deployed to evaluate the original query \(q\) alongside the detailed captions of the currently selected images, dynamically generating an adaptive sub-query \(d_k\) that explicitly specifies the missing narrative role required to advance bundle coherence (or issuing a DONE token to terminate expansion). Top-\(C = 5\) local candidate images are retrieved via \(d_k\), and each candidate branch is ranked using a dual-objective scoring function:
$\(\text{Score}(B_k) = \frac{1}{k} \sum_{i=1}^k \cos(e_{d_i}, e_{x_i}) + \lambda \cdot \cos\left(e_q, \frac{\sum_{i=1}^k e_{x_i}}{\|\sum_{i=1}^k e_{x_i}\|}\right)\)$
where \(e_{d_i}\) and \(e_{x_i}\) represent dense embeddings of the \(i\)-th sub-query and selected image, \(e_q\) is the global query embedding, and \(\lambda = 0.3\) regulates the trade-off between satisfying the instantaneous sub-query role and maintaining alignment with the macro-level narrative trajectory.
4. Whole-Bundle VLM Reranking: Overcoming Embedding Blindness via Holistic Structural Auditing
Although dense vector dot-products provide rapid heuristic guidance during beam exploration, low-dimensional embeddings fail to verify fine-grained relational logic, such as consistent personal identity across varying poses or subtle causal transitions between event phases. To guarantee the final compositional fidelity, BundleWeaver executes Whole-Bundle VLM Reranking. Completed candidate bundles from all surviving beam branches are pooled, and a high-capacity vision-language model (e.g., GPT-4o) evaluates each complete candidate bundle in a pointwise manner. Prompted with all constituent images simultaneously alongside the original query, the VLM conducts a rigorous structural audit—checking for complete stage coverage, identity preservation, and absence of irrelevant noise—and assigns an absolute coherence score from 1 to 10. The highest-scoring candidate is returned as the final bundle \(B^*\), effectively neutralizing the risk of fragmented compositions.
Benchmark & Evaluation Protocol¶
To benchmark IBC, the authors constructed IBCBench, comprising 109,467 images and 667 rigorously verified queries sourced from 57 YFCC-100M user archives with rich EXIF timestamps and GPS records (with user album identifiers discarded). The mining pipeline executed a five-stage filtration: (1) spatiotemporal session extraction yielding 7,460 candidate sliding windows scored by spatiotemporal coverage and semantic richness; (2) Claude-Opus-4.5 verification across two relational families (Same-location Dynamics, 52.5%; Cross-location Structures, 47.5%) with mandatory shared instance-level anchors; (3) a three-tier programmatic gate (Style, Anchor, and Template gates) that automatically rejected ungrounded candidates, accepting only ~8.5%; and (4) human review by 4 expert annotators equipped with an embedding-assisted nearest-neighbor panel to eliminate ambiguous or non-unique alternatives. The end-to-end acceptance rate was under 9%, producing bundles of 3 (24.3%), 4 (32.2%), and 5 images (43.5%). Evaluation measures set-level Precision, Recall, F1, and Exact Match (EM, the percentage of perfectly recovered ground-truth bundles) averaged per query.
Key Experimental Results¶
Main Results¶
The performance of BundleWeaver was evaluated against four representative baseline categories across all 667 queries in IBCBench: dense multimodal embedding models (CLIP, SigLIP2, Qwen3-VL-Embed, RzenEmbed), caption-based dense/sparse text retrieval (BM25, BGE-M3, Qwen3-Embedding), metadata-augmented heuristics, and agentic two-stage VLM decompose-and-rerank baselines (Qwen2.5/3-VL, Gemini-3-Flash, GPT-4o, Claude-Sonnet-4.5). For atomic rankers, oracle-size truncation (\(|B^*|\)) was granted to provide an upper-bound baseline setting.
| Method Category | Evaluated Model / Pipeline | Precision (%) | Recall (%) | F1 (%) | EM (%) |
|---|---|---|---|---|---|
| Multimodal Embedding | CLIP-ViT-B/32 | 2.21 | 2.21 | 2.21 | 0.00 |
| Multimodal Embedding | SigLIP2-giant | 9.97 | 9.97 | 9.97 | 0.15 |
| Multimodal Embedding | Qwen3-VL-Embed-8B | 12.14 | 12.14 | 12.14 | 0.00 |
| Multimodal Embedding | RzenEmbed-7B | 14.88 | 14.88 | 14.88 | 0.30 |
| Caption + Text Embedding | BM25 | 3.29 | 3.29 | 3.29 | 0.00 |
| Caption + Text Embedding | BGE-M3 | 7.05 | 7.05 | 7.05 | 0.15 |
| Caption + Text Embedding | Qwen3-Embedding-8B | 9.32 | 9.32 | 9.32 | 0.30 |
| Metadata Heuristics | Session Clustering | 4.84 | 4.84 | 4.84 | 0.60 |
| Metadata Heuristics | User + Session Clustering | 8.94 | 8.94 | 8.94 | 1.05 |
| Metadata Heuristics | Time Proximity | 6.67 | 6.67 | 6.67 | 0.45 |
| Metadata Heuristics | User + Spatiotemporal | 8.80 | 8.80 | 8.80 | 0.75 |
| VLM Decompose & Rerank | Qwen2.5-VL-72B | 11.39 | 10.16 | 10.67 | 0.15 |
| VLM Decompose & Rerank | Qwen3-VL-235B | 15.78 | 14.35 | 14.95 | 0.90 |
| VLM Decompose & Rerank | Gemini-3-Flash | 20.89 | 15.53 | 17.06 | 0.60 |
| VLM Decompose & Rerank | GPT-4o | 18.50 | 17.34 | 17.84 | 0.60 |
| VLM Decompose & Rerank | Claude-Sonnet-4.5 | 24.79 | 23.01 | 23.74 | 1.50 |
| Proposed Framework | BundleWeaver (Ours) | 30.95 | 30.46 | 30.28 | 7.20 |
| Relative Gain | vs. Strongest Baseline (Claude-Sonnet-4.5) | +24.8% | +32.4% | +27.5% | +380.0% |
Ablation Study¶
The ablation experiments isolate the operational contribution of each core component within BundleWeaver using the default GPT-4o backbone.
| Configuration | Precision (%) | Recall (%) | F1 (%) | EM (%) | Analysis & Architectural Impact |
|---|---|---|---|---|---|
| BundleWeaver (Full Model) | 30.95 | 30.46 | 30.28 | 7.20 | Full agentic hyperedge discovery & reranking pipeline |
| w/o Diverse Seed | 27.52 | 27.66 | 27.21 | 6.45 | Single-root search traps the agent in localized semantic clusters |
| w/o Contextual Candidate Pruning (CCP) | 24.57 | 24.74 | 24.69 | 5.85 | Unbounded global space introduces noisy, physically invalid distractors |
| w/o Beam Search (Greedy Search) | 25.63 | 25.93 | 25.35 | 6.30 | Premature greedy errors derail downstream narrative paths |
| w/o Whole-Bundle VLM Reranking | 26.33 | 26.70 | 26.13 | 6.15 | Embedding scores fail to catch subtle cross-image identity mismatches |
| Reference Baseline: GPT-4o Dec. & Re. | 18.50 | 17.34 | 17.84 | 0.60 | Independent sub-query retrieval without inter-image constraints |
Key Findings¶
- Relational Blindness in Static Decomposition: Even frontier VLMs like Claude-Sonnet-4.5 achieve a meager 1.50% EM under the decompose-and-rerank paradigm. This empirically validates Theorem 1: because sub-queries are matched independently against the pool, dominant distractors with high local visual alignment displace true bundle members, resulting in fatal identity and narrative breaks.
- CCP is Contextual Support, Not a Dataset Shortcut: Directly applying CCP heuristics to atomic models (e.g., RzenEmbed with CCP drops F1 from 14.88% to 8.80%) degrades retrieval quality, proving that spatiotemporal proximity alone does not imply narrative cohesion. In contrast, BundleWeaver without CCP still achieves 24.69% F1 and 5.85% EM, substantially outperforming the strongest GPT-4o baseline (17.84% F1, 0.60% EM), confirming that the framework's primary strength originates from adaptive role reasoning.
- Exploration vs. Noise Trade-off: Beam parameter sweeps exhibit a clear inverted U-curve peaking at beam width \(B_u = 4\) and candidate depth \(C = 5\). Over-expanding the search tree (\(B_u = 16, C = 15\)) floods the beam with marginally relevant candidates that distract the LLM's adaptive reasoning trajectory.
- Pointwise vs. Listwise Reranking: Pointwise VLM evaluation outperforms Listwise selection across F1 (30.28% vs. 29.94%) and EM (7.20% vs. 6.90%). Presenting dozens of candidate images simultaneously to a VLM induces attention dilution, whereas pointwise scoring forces an isolated, rigorous audit of each candidate bundle.
Highlights & Insights¶
- Formulation of Image Bundle Composition: Shifts the fundamental retrieval paradigm from atomic point-wise matching to the dynamic composition of non-decomposable, multi-image hyperedges.
- Theoretical Formalization of Relational Blindness: Provides a rigorous mathematical proof demonstrating that the IBC objective is strictly non-submodular, establishing the theoretical cutoff where independent top-\(k\) retrieval fails deterministically in the presence of dominant distractors.
- Rigorous Semi-Automated Benchmark Construction: IBCBench integrates multi-stage VLM gating with interactive nearest-neighbor human verification to ensure that every target bundle serves as an unambiguous, unique ground truth.
- Zero-Shot Agentic Versatility: BundleWeaver operates in a plug-and-play, training-free manner across multiple open-weight and proprietary backbones (Qwen3-VL, Gemini-3-Flash, GPT-4o, Claude-Sonnet-4.5), yielding consistent improvements across diverse model families.
Limitations & Future Work¶
- Reliance on Spatiotemporal Metadata: The current candidate pruning heuristic relies on valid EXIF timestamps and GPS tags. Adapting BundleWeaver to in-the-wild web collections or historical photo streams that lack physical metadata will require purely visual narrative continuity pruning.
- Inference Latency and Computational Overhead: Iterative LLM sub-query generation coupled with multi-image VLM reranking incurs an average latency of ~32 seconds per query on a 109K-image pool. Developing parameter-efficient fine-tuning (PEFT) methods for compact bundle-level scorers represents a critical path toward real-time deployment.
- Extension to Continuous Video Streams: While IBCBench focuses on discrete photo bundles, extending this compositional paradigm to continuous video stream extraction (Video Bundle Retrieval) introduces dense temporal dependencies and more complex causal dynamics.
Related Work & Insights¶
- vs. Traditional & Multi-Hop T2I Retrieval (e.g., MMEB, DeepImageSearch): Traditional and multi-hop retrieval systems evaluate relevance independently at the single-image level; IBC evaluates non-decomposable joint relevance across an entire image subset.
- vs. Composed Image Retrieval (e.g., CIRR, FashionIQ): CIR assumes a pre-existing query image paired with modifying text to find a single target image; IBC begins with text alone and must discover, filter, and assemble all visual assets from scratch.
- vs. Visual Storytelling (e.g., VIST, MMIU): Visual storytelling assumes that an ordered sequence of images is already provided and tasks the model with generating a textual narrative; IBC operates in the reverse direction, discovering the coherent image set from an unstructured pool given an abstract narrative query.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Formulates a pioneering retrieval paradigm, introduces the verified IBCBench benchmark, and frames composition as incremental hyperedge discovery.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive evaluation against four baseline paradigms, in-depth component ablations, hyperparameter sweeps, and formal theoretical proofs.
- Writing Quality: ⭐⭐⭐⭐⭐ Exceptionally articulate task formulation, lucid mathematical proofs of non-submodularity, and clean structural organization.
- Value: ⭐⭐⭐⭐⭐ Bridges the critical gap between atomic image ranking and human episodic memory retrieval, providing a strong baseline for future agentic visual systems.