Mosaic Chunk gives AI video generation long-term memory

Mosaic Chunk is an open-source memory system built to solve one of the most frustrating flaws in AI video: objects morphing or disappearing the moment they leave the frame.

Most video generators struggle with temporal consistency. Once an actor or a vehicle exits the shot for even a couple of seconds, the model effectively gets amnesia and redraws them differently. Mosaic Chunk gives existing pipelines a way to remember.

What it does

Instead of trying to store every single previous frame—which would quickly blow up your VRAM—Mosaic Chunk selectively tracks key visual features from earlier in the clip.

It identifies the essential pieces of information about objects and environments, tucks them into memory, and retrieves them whenever those elements re-enter the scene.

In early demos, the system successfully maintains an object's appearance and keeps the broader environment consistent, even after a target has been hidden or out of view for several seconds.

How to run it

The team released both the code and pre-trained weights for local execution, along with full setup instructions on the project page.

There are two separate models depending on your workflow. The text-to-video version is built on top of MiniMax H3, while the image-to-video release is powered by Ling World Infinity.

You can head to the repository, scroll down to the setup guide, and follow the steps to run the pipeline locally on your own machine.

Why it matters

Consistency across cuts has been the missing piece holding AI generation back from real filmmaking. Tacking on selective memory rather than brute-forcing longer context windows is a smart, resource-efficient way to tackle temporal drift.

If you are an indie creator or researcher already experimenting with open-source video models, this is well worth spinning up locally to see if it fixes your continuity headaches.