Breaking

ByteDance and NTU release StoryMem for memory-based long video generation

July 3, 2026 at 09:29 EDT

Media Generation
Research & Papers
Open Source

A research team from ByteDance and Nanyang Technological University (NTU) has released StoryMem, a framework that generates multi-shot long video stories using an explicit visual memory. It extends pre-trained single-shot video diffusion models into roughly one-minute continuous narrative clips while keeping characters and scenes consistent.

December 2025 · NTU Singapore × ByteDance

StoryMem: a minute-long AI video that keeps its characters looking the same

An "explicit visual memory" bolted onto a pretrained single-shot diffusion model lets StoryMem build multi-shot narrative videos shot by shot — holding a character's appearance stable across cuts. Code, weights and a new benchmark are open-sourced.

+28.7%

Cross-shot consistency vs. base Wan2.2

~1 min

Multi-shot narrative video length

LoRA

Fine-tuning only — small added cost

Cross-shot consistency, drawn to scale

Relative consistency score — taller column is more stable across cuts

100

Base
Wan2.2

→

128.7

StoryMem
(+28.7%)

Memory-to-Video (M2V): consistency shot by shot

Generate shot

Produce the next shot with the Wan2.2 Video DiT

→

Store keyframes

3D-VAE encode into a compact memory bank

→

Condition next shot

Inject memory via latent concat + negative RoPE shift

Where it shines

Strong character consistency and smooth shot transitions
Ahead of HoloCine, StoryDiffusion, IC-LoRA and Sora2 in samples
Fit for previsualization, storyboarding and short films

Open questions

Memory-bank size and keyframe selection affect long-range consistency
Very long videos still need further optimization
Open release invites community fine-tuning and integration

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.