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
EMNLP'26 Main Conference
cs.CV, cs.IR
2026-08-27
IBC composes relational photo bundles from 109k images. Best embedder ~15 F1; best decompose-and-rerank 23.74 F1 / 1.50% EM. BundleWeaver reaches 30.28 F1 and 7.20% EM.
Text-to-image retrieval still scores each photo alone. From MSCOCO and Flickr30K through later multi-hop visual search, the target remains a ranked list of disconnected images. Queries over a personal library are often stories: the Eiffel Tower from day to night, the beats of a concert. The right object is a small bundle tied by time, event, or space. Two independently great daytime shots can still fail a transition query.
Image Bundle Composition makes that the task. Bundles are not pre-indexed; any subset could be an answer, so the search space scales like N to the power Kmax. Joint relevance is not a monotone sum of per-image scores.
IBCBench keeps 109,467 YFCC-100M photos with spatiotemporal metadata, mines sliding windows inside user sessions, verifies relations with Claude Opus 4.5, then has four annotators drop ambiguous cases. The final set has 667 queries and an end-to-end accept rate under 9%. Bundle sizes 3, 4, and 5 are 24.3%, 32.2%, and 43.5%. Same-location dynamics are 52.5%, cross-location structure 47.5%. A valid bundle must be jointly complete, cross-image bound, globally as unique as possible, and internally non-redundant.
BundleWeaver treats IBC as query-conditioned hyperedge discovery. An LLM extracts a visual anchor and diverse seeds so the search does not collapse onto near-duplicates. Each branch grows by beam search: the LLM writes a sub-query for the missing role, candidates are drawn from a spatiotemporal neighborhood of the seed, and a score mixes step-wise match with whole-bundle alignment to the original query. Completed paths go to a VLM that rates structural coherence from 1 to 10. Depth is capped; subsets are never enumerated.
Pointwise multimodal embeddings barely move. CLIP-ViT-B/32 scores 2.21 F1 and 0 exact match; the strongest, RzenEmbed, reaches 14.88 F1 and 0.30 EM. Caption-then-text retrieval and session or spatiotemporal reranking stay low; rigid spatiotemporal filters can even hurt F1, so proximity is not a substitute for filling roles.
The strongest baseline family is static decompose-then-rerank. Claude Sonnet 4.5 reaches 23.74 F1 and 1.50% EM; GPT-4o reaches 17.84 F1 and 0.60% EM. Decomposition helps recall. Exact match stays poor because sub-queries never constrain each other.
BundleWeaver scores 30.28 F1 and 7.20% EM. Exact match is about 3.8× the best baseline; Precision, Recall, and F1 rise 24.8%, 32.4%, and 27.5%. Dropping diverse seeds, candidate pruning, beam search, or VLM whole-bundle rerank yields 27.21, 24.69, 25.35, and 26.13 F1, each still above the GPT-4o decompose baseline. Swapping in Qwen3-VL-235B, Gemini 3 Flash, Claude, or GPT-4o as backbone jumps EM from below 2% to 4.35% through 7.20%. Larger bundles hurt every method; incremental expansion drops less.
A CLIP top-k over a camera roll will keep returning photos that each look relevant and do not form the story. IBC changes the unit of evaluation to a set: exact match means the system recovered that episode. A training-free agent is enough to move EM from about 1.5% to 7.2%, by filling missing roles and checking the whole bundle, not by swapping in a larger embedder. 7.2% is still far from a product, but it is no longer the atomic ranking problem.
All 667 queries come from YFCC personal photos. Medical sequences, legal evidence, and continuous video are out of scope. Bundles are length 3 to 5 from sliding windows; real K and relation types will be wilder. Annotation is anchored on spatiotemporal sessions, and search-time pruning uses the same neighborhood, so there is a structural leak. The ablation still beats decompose-and-rerank without pruning. The method calls frontier VLMs zero-shot; cost and latency are not first-class metrics. 7.2% EM also says the task is brutally hard, and the benchmark may over-prefer a single unique bundle where a real library would accept several. There is no end-to-end fine-tune.