New LongBanana benchmark evaluates agent-driven long-context image synthesis

Researchers have introduced LongBanana, a novel benchmark designed for evaluating long-context multi-reference image synthesis, particularly in the context of agent-driven image generation. This benchmark requires systems to retrieve dispersed visual information, disregard irrelevant data, and accurately bind entities and relations to compose a single coherent image from multiple sources. LongBanana features 90 tasks, each with numerous reference images and constraints, resulting in over 4,670 expert-specified evaluation items. The system also includes LongBanana-Gen, a framework that utilizes agent harnesses to improve image generation performance, with a ReAct-Plan-Exec-Reflect workflow showing particular gains in evidence grounding and repair. AI

IMPACT Establishes a new evaluation standard for complex, multi-reference image generation tasks, potentially driving advancements in agent-based AI systems.

RANK_REASON The cluster describes a new academic benchmark and associated framework for image synthesis, detailed in an arXiv paper. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• LongBanana
• LongBanana-Gen
• ReAct-Plan-Exec-Reflect
• ScienceCast

• paper
• product

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →