Research map / Vision-language and multimodal / Multimodal LLMs

LogiStory: A Logic-Aware Framework for Multi-Image Story Visualization

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 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.

Proposes LogiStory, a multi-agent planner plus causal verification and refinement loop for logic-aware multi-image story visualization, and introduces the LogicTale benchmark with causal annotations.

Problem

Multi-image story visualization often lacks logical flow: actions, states, and characters become disjointed, so generated visual narratives are hard to follow even when individual images look good.

Approach

Model "visual logic" explicitly: a multi-agent planner (SceneCrafter, LogicMiner, ShotPlanner) extracts entities and causal event tuples, and a Visual Logic Enhancement Module uses a local MLLM causal monitor and a global causal graph verifier to refine panels with image editing tools.

Key results

• LogiStory scores 4.45 Narrative Causality and 0.8267 Story Readability on LogicTale, the top values in the automatic table, versus 4.08 and 0.7440 for the next-best Nano Banana.
• Instance Consistency is 4.23 for LogiStory, behind GPT-4o + GPT-image-1 at 4.67 on the automatic metric.
• Replacing the multi-agent planner with DeepSeek-V3 planning drops Narrative Causality from 4.45 to 3.64 in the planning ablation.

Surprising findings

• Adding only the Global Causal Verifier or only the Local Causal Monitor gives smaller gains than using both together, and the Local Causal Monitor alone lowers Narrative Causality relative to the verifier-only variant.
• Smaller VLM backbones such as Qwen2.5-VL-7B yield lower logic scores, and the authors attribute this partly to smaller models assigning uniformly high causal scores that fail to trigger refinement.

Limitations

• Evaluation relies heavily on MLLM and LLM judges and a small 60-story benchmark.
• The authors note that the bottleneck on complex stories is the generation side, with subtle emotional expression and event-level coherence sometimes degrading.
• Code and dataset are promised only 'in the near future', and a subset is provided in supplementary material.

Experimental setup

Datasets: LogicTale ViStoryBench StoryDiffusion stories

Benchmarks: LogicTale ViStoryBench-Lite

Metrics: Instance Consistency Narrative Causality Story Readability Aesthetic Quality (HPSv2) Style Consistency (DINOv2) Character Expressiveness Krippendorff's Alpha Pearson correlation Kendall's Tau

Baselines: Nano Banana Gemini 2.0 Flash GPT-4o + GPT-image-1 MM-StoryAgent StoryDiffusion ConsiStory SDXL Flux StoryGen Story-Adapter

Builds on: Flux DeepSeek-V3 StoryDiffusion ConsiStory MM-StoryAgent

Closest papers by meaning

• ViStoryBench: Comprehensive Benchmark Suite for Story Visualization (CVPR 2026 · similarity 0.89)
• StoryGPT-V: Large Language Models as Consistent Story Visualizers (CVPR 2025 · similarity 0.88)
• VinaBench: Benchmark for Faithful and Consistent Visual Narratives (CVPR 2025 · similarity 0.88)
• SceneDecorator: Towards Scene-Oriented Story Generation with Scene Planning and Scene Consistency (NeurIPS 2025 · similarity 0.87)
• Factuality Matters: When Image Generation and Editing Meet Structured Visuals (ICLR 2026 · similarity 0.87)
• Intelligent Grimm - Open-ended Visual Storytelling via Latent Diffusion Models (CVPR 2024 · similarity 0.87)