LongBanana benchmark measures image creation from many references
LongBanana: An Expert-Verified Benchmark for Long-Context Multi-Reference Image Synthesis
Computer Vision and Pattern RecognitionComputation and Language
Summary
Creating one clear picture from many different images is very hard, because you have to pick the right parts, ignore distractions, and put everything together correctly. The authors made LongBanana, a new test with 90 special tasks to check how well computers do this. Each task gives many images and rules, and experts carefully marked exactly what the final picture should show. They also made a system to help computers improve by thinking and fixing mistakes step by step. This work helps make better tools for making pictures using lots of information.
What this means in practice
- •For image generation developers: Evaluate and improve software that generates images from many source pictures using expert-verified benchmarks and structured tests.
- •For ai workflow engineers: Implement agent-based workflows that plan and fix multi-image generation tasks to increase output quality.
Authors
Haoxiang Cao, Yuxuan Zhang, Penghui Du, Bo Li, Jiajiong Cao, Yingying Fan, Zimeng Wu, Yifu Luo, Mingzhe Zheng, Junwen Miao, Huaisong Zhang, Yang Yao, Yu Wang, Dongfu Jiang, Ping Nie, Wenhu Chen, Chaoqun Wang, Changqian Yu, Kelsey R. Allen
Abstract
We introduce LongBanana, to our knowledge the first expert-designed and expert-verified benchmark for long-context multi-reference image synthesis in the age of agent-driven image generation. The task requires retrieving dispersed visual evidence, rejecting distractors, binding entities and relations, and composing one coherent final image from multiple references. LongBanana contains 90 carefully designed tasks organized by a 3-by-3 taxonomy; each task includes about 32 reference images, 20 constraints, and 50 evaluation items, yielding 4,670 expert-specified atomic evaluation items in total. Its graph-grounded evaluation design instantiates each task with reference images and an entity-relation graph, then converts node, edge, and whole-graph obligations into checklist items for Content Fidelity, Relational Binding, and Global Constraint Satisfaction. Reviewed expert labels include a 15% double-labeled subset with 89.4% exact agreement and 0.83 quadratic weighted kappa. We further introduce LongBanana-Gen, an agent-harness framework for long-horizon image generation; experiments show that minimal agent harnesses improve over direct generation, while a ReAct-Plan-Exec-Reflect (RPR) workflow yields harness-dependent gains by strengthening evidence grounding and repair. Together, LongBanana provides an auditable testbed for measuring dense visual evidence use and a shared target for developing stronger agent-driven image-generation systems. Code and data are available at the project page: https://long-banana.com/.