Research map / Vision-language and multimodal / Multimodal LLMs

Narrative Weaver: Towards Controllable Long-Range Visual Consistency with Multi-Modal Conditioning

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.

Narrative Weaver pairs an MLLM narrative planner (Qwen2.5-VL-3B) with a Flux.1-Dev diffusion generator via learnable queries and a decaying-length memory bank, trained in three progressive stages, and releases the EAVSD e-commerce storyboard dataset.

Problem

Generating long-range, narratively coherent visual sequences with visual consistency and fine-grained control from multimodal (image + text) inputs; existing methods rely on text-only conditioning or lack consistency across frames, and no suitable dataset exists.

Approach

A hybrid AR (MLLM) + diffusion design where the MLLM plans text narratives and compresses history into learnable queries, combined with VAE features of the conditioning image and a memory bank of pooled prior-frame VAE features with geometric length decay, trained with a three-stage progressive schedule and a dynamic attention mask.

Key results

• On the LLM-judged consistency evaluation, Narrative Weaver scores 7.54 ITC and 8.67 MSSC, the best among compared methods, but Flux.1-Kontext scores higher on RGC (9.41 vs 8.86).
• Automated DreamSim story-level score of 12.18 is lowest (best) among methods except Flux.1-kontext (3.71), which the authors attribute to copy-paste behavior; Narrative Weaver's CLIP score is 89.98 vs 97.17 for Flux.1-kontext.
• Ablation: stage 2 plus stage 3 yields average 8.10 vs 7.12 without either stage.

Surprising findings

• Flux.1-Kontext achieves the highest automated CLIP Score and lowest DreamSim, yet the authors attribute this to copy-paste behavior rather than true narrative consistency.
• Narrative Weaver does not beat SEED-Llama-14B on the IRS metric (2.77 vs 3.28) in the CoMM narrative planning evaluation.

Limitations

• Limited to image generation; video extension left to future work.
• Dataset covers only the e-commerce domain.
• Automated metrics (CLIP, DreamSim) reward frame similarity and favor copy-paste baselines, so user study is needed.

Experimental setup

Datasets: EAVSD OmniGen2 CoMM CI-VID 30M public text-image pairs

Benchmarks: EAVSD OmniGen2 test subset CoMM CI-VID

Metrics: GPT-4o score (ITC, RGC, MSSC, MSCC, IMQ) CLIP Score DreamSim Style Entity Trend CPL ImgQ IRS user preference

Baselines: TALC StoryDiffusion IP-Adapter AnimeShooter Flux.1-Kontext Qwen-Image-Edit MiniGPT-5 SEED-Llama-8B SEED-Llama-14B

Builds on: Flux.1-Dev Qwen2.5-VL FlexAttention StoryDiffusion

Closest papers by meaning

• SceneDecorator: Towards Scene-Oriented Story Generation with Scene Planning and Scene Consistency (NeurIPS 2025 · similarity 0.88)
• Intelligent Grimm - Open-ended Visual Storytelling via Latent Diffusion Models (CVPR 2024 · similarity 0.86)
• Storybooth: Training-Free Multi-Subject Consistency for Improved Visual Storytelling (ICLR 2025 · similarity 0.86)
• VinaBench: Benchmark for Faithful and Consistent Visual Narratives (CVPR 2025 · similarity 0.86)
• LogiStory: A Logic-Aware Framework for Multi-Image Story Visualization (ICLR 2026 · similarity 0.85)
• StoryDiffusion: Consistent Self-Attention for Long-Range Image and Video Generation (NeurIPS 2024 · similarity 0.85)