KlingAI releases LongBanana: 90-task benchmark for multi-reference image synthesis
_akhaliq · x · 2026-11-10
KlingAI Research released LongBanana, an expert-verified benchmark for long-context multi-reference image synthesis: 90 tasks, 32 references per case, 4,670 atomic evaluation items.
Per the paper figure, tasks ask models to render complex scenes (e.g. a boutique shelf) while honoring dozens of reference constraints ("structured black rectangular bag must exist", "authentic leather texture must not exist", etc.), scored 0–10 by an Agentic Rubric Scorer across Coverage, Relation, Appearance and Reasoning, with tool-assisted verification (OCR, counting, layout matching) — targeting current models' failures on many-reference, long-instruction generation.
More from Multimodal
- AI-built 3D desert game world editor demo stuns with "They can build now?" — jasonkneen · 2026-11-10
- Open-source full-duplex voice model Sophon-Realtime v1 beats Grok on tool calling — bdsqlsz · 2026-11-10
- MiniMax H3-Image enters early access with 16-reference-image precision — SarahAnnabels · 2026-11-10
- Disney partners with Adobe Firefly to build brand-safe custom AI tools — The Decoder · 2026-11-10
- GPT-image 2.6 'brandshaped typography' prompt turns brand text into logo shapes — aziz4ai · 2026-11-10
- SplatAnything debuts: feed-forward 3DGS reconstruction that generalizes to arbitrary scenes — _akhaliq · 2026-11-10