Research map / Image generation / Autoregressive and latent generation
LongBanana: An Expert-Verified Benchmark for Long-Context Multi-Reference Image Synthesis
Haoxiang Cao, Yuxuan Zhang, Penghui Du, Bo Li, Jiajiong Cao, Yingying Fan, Zimeng Wu, Yifu Luo, Mingzhe Zheng, Junwen Miao, Huaisong Zhang, YANG YAO and 7 more
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces LongBanana, an expert-designed and expert-verified benchmark for long-context multi-reference image synthesis under agent-controlled image generation.
Problem
Evaluating image synthesis that must retrieve dispersed visual evidence from long multi-reference contexts, reject distractors, and bind entities and relations into one coherent image.
Approach
A benchmark of expert-designed, expert-verified tasks for long-context multi-reference image synthesis in the setting of agent-controlled image generation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| decompose prompt with privileged graph (PD-Direct) instead of full-prompt direct generation (ablation) | Nano Banana 2 · LongBanana | overall average: 68.4 -> 74.6 | Improved |
| use OpenClaw + RPR agent instead of PD-Direct | Nano Banana 2 · LongBanana | P@70%: 35.6 -> 47.8 | Improved |
| add RPR workflow to OpenClaw harness (ablation) | OpenClaw · LongBanana | overall average: 76.1 -> 78.4 | Improved |
| add RPR workflow to OpenCode harness (ablation) | OpenCode · LongBanana | P@70%: 47.8 -> 43.3 | Hurt |
| add RPR workflow to Hermes-Agent harness (ablation) | Hermes-Agent · LongBanana | P@70%: 43.3 -> 35.6 | Hurt |
| add RPR workflow to Hermes-Agent harness (ablation) | Hermes-Agent · LongBanana | overall average: 73.9 -> 75.6 | Improved |
| add RPR to OpenClaw on temporal/state reasoning (ablation) | OpenClaw · LongBanana TSR | checklist average: 78.7 -> 73.6 | Hurt |
| ReAct only, Plan-Exec only, Reflect only vs full RPR (ablation) | OpenClaw · LongBanana | overall average and pass rate | Improved |
| use Nano Banana 2 instead of Seedream 4 or GPT Image 1 backend | OpenClaw + RPR · LongBanana | overall average: 63.6 -> 78.4 | Improved |
| use agentic checklist judge instead of single-pass VLM-as-judge | Gemini 3 Flash judge · audit pool of 900 outputs | Pearson correlation with human: 0.8413 -> 0.8707 | Improved |
| use agentic checklist judge instead of VLM-as-judge | Gemini 3 Flash judge · item level | opposite-end 0<->2 errors: 44.09% -> 24.81% | Improved |
Experimental setup
Benchmarks: LongBanana
Closest papers by meaning
- ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both (NeurIPS 2026 · similarity 0.88)
- AgentVista: Evaluating Multimodal Agent in Ultra-Challenging Realistic Visual Scenarios (NeurIPS 2026 · similarity 0.87)
- VisionDirector: Vision-Language Guided Closed-Loop Refinement for Generative Image Synthesis (CVPR 2026 · similarity 0.87)
- GIR-Bench: Versatile Benchmark for Generating Images with Reasoning (ICLR 2026 · similarity 0.87)
- MMMG: A Massive, Multidisciplinary, Multi-Tier Generation Benchmark for Text-to-Image Reasoning (NeurIPS 2025 · similarity 0.87)
- Hummingbird: High Fidelity Image Generation via Multimodal Context Alignment (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
