Editorial illustration for Microsoft Research Mirage adds persistent spatial memory to video generation
Microsoft Research Mirage adds persistent spatial memory...
Video generation has long suffered from a quiet, expensive flaw: every time a model renders a new viewpoint, it must rebuild the world from scratch, pixel by pixel, memory bleeding away with each frame. Microsoft Research’s Mirage refuses to play that game. Instead of hoarding visible color points, a “double bottleneck” that guzzles compute and leaks information, it keeps the model’s own internal features, each pinned to a precise spot in 3D space.
That store becomes a persistent spatial memory. When the camera moves, the system simply projects this cache onto the new view and feeds it directly to the generator. No point cloud rendering.
No wasteful re-encoding. The trick is that the data lives at the model’s compact internal resolution, slashing memory use. And Mirage grows that memory segment by segment: seeding from the first image, reading what’s relevant, generating fresh frames, then writing back only stable geometry, filtering out moving objects and sky before they corrupt the long-term archive.
Built on Alibaba’s Wan2.2 with a lightweight add-on and fine-tuned with LoRA, it leaves pixel-based rivals like Spatia and general generators like Wan2.1 far behind on the WorldScore benchmark. Mirage doesn’t just remember what’s around the corner. It remembers without wasting a single pixel.
Mirage is a new video world model that skips the costly detour through pixel-based memory. That speeds up generation and keeps a scene's spatial structure stable even during long camera moves.
Mirage doesn’t just patch a leak. It rewires the pipeline. By storing what the model already knows, its own internal features, Microsoft’s team sidesteps the expensive, lossy loop of rendering and re-encoding.
The result is a system that remembers where things are, not just what they looked like. That distinction matters. It means coherent video that doesn’t hallucinate a new room every time the camera pans.
It means memory that grows with the scene, not against it. The benchmark scores confirm it: Mirage is faster, lighter, and more spatially aware than its peers. This isn’t a minor optimization.
It’s a structural shift. Video generation has long suffered from amnesia. Mirage gives it a place to keep its notes.
Related Reading
Browse all Microsoft news →
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Databricks finds multi-step agents beat single-turn RAG by 21% to 38% on STaRK
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
Related Topic Claude gains shared context in Excel, PowerPoint; Microsoft adds Copilot Cowork
AI Tools & Apps
Related Topic Microsoft adds ’vibe working’ to Word and Excel; Copilot Agent Mode now default
Business & Startups
From Archives Amazon security research prompts White House ban on Anthropic Fable
From Archives Study: AI coding agents locate correct file but miss key lines in bugs
From Archives Microsoft and OpenAI split; both prepare for legal battle over AI training
Business & Startups
From Archives Microsoft unveils MXC OS-level sandbox for AI agents, OpenAI, Nvidia join
Business & Startups
Common Questions Answered
How does Microsoft Research's Mirage solve the persistent spatial memory problem in video generation?
Mirage stores the model's internal features pinned to precise locations in 3D space rather than rebuilding the world from scratch for each new viewpoint. This persistent spatial memory approach eliminates the expensive and lossy process of rendering and re-encoding pixels, allowing the system to maintain coherent information across frames without information loss.
What is the 'double bottleneck' problem that Mirage addresses in video generation?
The double bottleneck refers to the traditional approach of hoarding visible color points, which consumes excessive computational resources and causes information to leak away with each frame. Mirage sidesteps this inefficient pipeline by storing internal features instead of relying on rendered color data, significantly reducing computational overhead.
Why does storing internal features instead of color points improve video coherence in Mirage?
By remembering where things are located in 3D space rather than just what they looked like visually, Mirage prevents the model from hallucinating new details when the camera pans or changes viewpoint. This spatial awareness allows the system to generate coherent video sequences where objects maintain consistent positions and appearances throughout the scene.
How does Mirage's memory system scale differently compared to traditional video generation models?
Mirage's memory grows with the scene rather than against it, meaning the spatial memory becomes more useful as the scene expands rather than becoming a bottleneck. This scalability improvement allows the model to handle increasingly complex scenes without the performance degradation seen in traditional approaches that rebuild information with each frame.
Further Reading
- Latent Spatial Memory for Video World Models — arXiv
- Mirage | Latent Spatial Memory for Video World Models — Microsoft Research
- Latent Spatial Memory for Video World Models — Hugging Face Papers
- Spatia: Video Generation with Updatable Spatial Memory — Microsoft Research
- Neural World Simulators with Persistent 3D State — Microsoft Research
