Research map / Vision-language and multimodal / Multimodal LLMs

ViStoryBench: Comprehensive Benchmark Suite for Story Visualization

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 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 ViStoryBench, a benchmark of 80 multi-shot stories with 344 characters and 509 reference images across 10 styles, plus 12 automated metrics validated by human study, used to evaluate about 25 story visualization methods.

Problem

Existing story visualization benchmarks are narrow: short prompts, no character references, or single-image cases, which fail to reflect real-world narrative complexity and obscure true model performance.

Approach

A reference-based, multi-shot story benchmark with structured shot scripts (setting, plot, onstage characters, static shot, camera perspective) built from curated stories with LLM-assisted scripting and human review, paired with a hybrid evaluation suite mixing expert feature models (Grounding DINO, CLIP/face models, CSD) and VLM judges (Gemini-3-Pro, Qwen3-VL).

Key results

• Copy-Paste Baseline achieves best scores on many consistency metrics but very low prompt alignment (Scene 0.34), showing metrics must be combined.
• Gemini-3.0 Pro achieves the top average alignment (3.37) on ViStoryBench-Lite among evaluated methods, while GPT-4o ranks 2nd (3.34).
• Automated Self-CIDS correlates with human character consistency judgments at Spearman 0.6759 and Pearson 0.7956; aesthetics correlates weakest (Kendall 0.2565).

Surprising findings

• The trivial copy-paste baseline scores highest on several consistency and aesthetic metrics, exposing the need for copy-paste and alignment metrics.
• Multi-modal LLMs like GPT-4o excel at narrative alignment but lag in Inception Score and aesthetics.

Limitations

• Focuses only on multi-image consistency; no audio or video temporal evaluation.
• No background reference images or scene-level image similarity evaluation.
• VLM-based scoring is prone to hallucination and expert models may underperform in complex contexts.

Experimental setup

Datasets: ViStoryBench

Benchmarks: ViStoryBench ViStoryBench-Lite

Metrics: CSD CIDS Prompt Alignment (Scene, Shot, CI, IA) OCCM Inception Score Aesthetic score Copy-Paste Rate

Baselines: StoryGen TheaterGen StoryDiffusion SEED-Story Story-Adapter UNO OmniGen2 CharaConsist QwenImageEdit-2509 GPT-4o Gemini-2.5 Gemini-3.0 Pro

Builds on: StoryDiffusion CSD Grounding DINO Aesthetic Predictor V2.5

Closest papers by meaning

• StoryGPT-V: Large Language Models as Consistent Story Visualizers (CVPR 2025 · similarity 0.91)
• LogiStory: A Logic-Aware Framework for Multi-Image Story Visualization (ICLR 2026 · similarity 0.89)
• VinaBench: Benchmark for Faithful and Consistent Visual Narratives (CVPR 2025 · similarity 0.89)
• Intelligent Grimm - Open-ended Visual Storytelling via Latent Diffusion Models (CVPR 2024 · similarity 0.88)
• ContextualStory: Consistent Visual Storytelling with Spatially-Enhanced and Storyline Context (AAAI 2025 · similarity 0.88)
• StoryWeaver: A Unified World Model for Knowledge-Enhanced Story Character Customization (AAAI 2025 · similarity 0.87)