Research map / Video / Video generation
Advancing Narrative Long Video Generation via Training-Free Identity-Aware Memory
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 IAMFlow, a training-free identity-aware memory framework for autoregressive narrative long video generation that organizes memory around persistent entity IDs, plus an accelerated inference pipeline and the NarraStream-Bench benchmark.
Problem
Autoregressive video generators lose entity identities and attributes over long, multi-prompt narratives because their memory (sink tokens, frame-level retrieval) is not organized around persistent entities.
Approach
Builds an ID-centric memory bank: an LLM assigns persistent global IDs to entities from prompts, evicted chunks are scored by text-key/visual-key compatibility and VLM fidelity, and greedy set-cover retrieval injects KV caches of frames covering active identities; adds adaptive cosine-blended prompt transitions and async VAE/VLM verification, all without training.
Key results
- On NarraStream-Bench, IAMFlow reaches an overall score of 75.73 versus 73.17 for LongLive and 72.88 for MemFlow, with Boundary Smoothness 65.03 vs 49.19 (LongLive).
- End-to-end latency on H20 GPUs is 156.69 s vs 264.86 s for Self Forcing, a 1.69x speedup, and FPS 6.20 vs 6.14 for LongLive.
- IAMFlow is best on five of six VBench-Long metrics against LongLive and MemFlow.
Surprising findings
- Removing the identity-aware memory entirely drops overall score most (75.73 to 74.28 in ablation), but the VLM-verification and LLM-planning removals cost comparatively little.
Limitations
- Conservative when prompts demand abrupt semantic changes, complex actions, or fine-grained relational updates.
- Entity handling restricted to human characters in the extraction prompt.
- Identity memory relies on LLM/VLM calls, adding dependence on auxiliary models.
Experimental setup
Datasets: NarraStream-Bench prompt database
Benchmarks: NarraStream-Bench VBench-Long
Metrics: Subject Consistency Background Consistency Temporal Flickering Motion Smoothness VTSS Boundary Smoothness Conditional Adjacent Consistency Conditional Long-range Consistency Entity Grounding Dynamic Trajectory
Baselines: Self Forcing Rolling Forcing Deep Forcing LongLive MemFlow
Builds on: MemFlow LongLive Self Forcing StreamingLLM
Closest papers by meaning
- EntityBench: Towards Entity-Consistent Long-Range Multi-Shot Video Generation (NeurIPS 2026 · similarity 0.89)
- One-Prompt-One-Story: Free-Lunch Consistent Text-to-Image Generation Using a Single Prompt (ICLR 2025 · similarity 0.87)
- MovieDreamer: Hierarchical Generation for Coherent Long Visual Sequences (ICLR 2025 · similarity 0.86)
- ARGUS: Stacked Multi-View Identity Mosaic Injection for Subject-Preserving Video Generation (NeurIPS 2026 · similarity 0.86)
- StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text (CVPR 2025 · similarity 0.86)
- Storynizor: Consistent Story Generation via Inter-Frame Synchronized and Shuffled ID Injection (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.