LongBanana: An Expert-Verified Benchmark for Long-Context Multi-Reference Image Synthesis
Kwai-Kolors · hf · 2026-11-09
Kwai-Kolors and academic collaborators introduce LongBanana, claimed to be the first expert-designed and verified benchmark for long-context multi-reference image synthesis in the agent-driven era. Tasks require retrieving dispersed visual evidence, rejecting distractors, binding entities and relations, and composing one coherent image from 32 references each. It contains 90 tasks in a 3×3 taxonomy with 4,670 expert-specified atomic evaluation items; a 15% double-labeled subset shows 89.4% exact agreement and 0.83 quadratic weighted kappa. Experiments show minimal agent harnesses beat direct generation, with a ReAct-Plan-Exec-Reflect workflow yielding further gains. Code and data are open.
More from Multimodal
- AI-generated "Two of Fucks" tarot card meme goes viral — MickeySteamboat · 2026-11-09
- Dynamic Gated Pathways wire every ViT layer to every LLM layer, boosting MLLMs over LLaVA-1.5 — HKUST · 2026-11-09
- AI-made Japanese horror anime "THE SECRET MASKS" tests AI storytelling — IntentionSubject2843 · 2026-11-09
- AnimaEdit LoRA brings reference and instruction-based image editing to Anima — Intrepid_Swim_8669 · 2026-11-09
- New Midjourney style code --sref 2219508089 yields adorable consistent troll series — michaelrabone · 2026-11-09
- ByteDance Seed launches Seedream 5.5: 100-round stable editing and native RGBA transparency — lyraxana · 2026-11-09