AI video generation has a memory problem. Every time a camera pans away and comes back, the scene changes. Furniture shifts. Textures flicker. Walls move. The model forgot what the room looked like because it never stored it anywhere.
Microsoft Research just released Mirage, a video world model that fixes this by giving AI a spatial memory that actually works. Instead of storing what a scene looks like as coloured pixels and converting them back and forth, Mirage remembers scenes in the model’s own internal language. The result: videos that stay geometrically consistent even when the camera takes long detours, generated up to 10 times faster and using 55 times less memory than previous approaches.
What Mirage Actually Does
Give Mirage a single image of a room and a camera path, and it generates a video that follows that path through the space. The camera can swing around corners, loop back to where it started, and traverse long trajectories. The room stays the room. The furniture stays where it was. The walls do not morph.
This is harder than it sounds. Existing video world models like Spatia, Voyager, and WonderWorld all try to maintain spatial consistency, but they do it by building a 3D point cloud of the scene in RGB (coloured pixels), rendering that cloud into an image at every new camera position, and then re-encoding that image back into the model’s internal format. Every single frame requires this round trip: render the cloud, re-encode, feed to the generator. It is slow, it eats GPU memory, and the pixel conversion loses information each time.
Mirage skips that conversion entirely. It stores the model’s own latent features (the compressed internal representation it already works with) directly at 3D locations in the scene. When the camera moves, Mirage projects this latent memory onto the new viewpoint at the model’s native resolution, no pixel rendering needed. The generator reads from memory and writes to memory, all in its own language.
Why This Matters for Creative Work
If you generate video, you have felt this problem. You prompt for a walk-through of an interior, and by the time the camera circles back, the room has changed. The couch is now a table. The window is a door. This inconsistency makes AI-generated spatial video useless for anything that requires continuity, which is most narrative and commercial work.
Mirage’s spatial memory means:
- Architectural walkthroughs that actually maintain the building’s layout across long camera moves
- Real estate virtual tours where rooms do not reshape themselves mid-video
- Game environment prototyping where an AI-generated space stays consistent as you explore it
- Film previsualisation where camera moves through generated sets stay spatially coherent
- Product walkthroughs where the environment around a product does not morph and shift
The efficiency gains are equally important. Ten times faster generation and 55 times less memory means you can actually run longer trajectories without crashing your GPU or waiting hours. Previous RGB-cache systems got progressively slower and more memory-hungry as the scene grew. Mirage’s per-frame cost stays nearly flat.
The Numbers That Matter
Mirage hits a WorldScore average of 70.36, beating the previous best spatial memory system (Spatia at 69.73) and leaving all general video generators (Wan2.1, CogVideoX, etc.) far behind. The strongest results are on 3D consistency (92.21 vs Spatia’s 86.40) and photographic consistency (93.95 vs 89.10), which is exactly where you would expect latent memory to help. The model remembers not just colours but semantic structure, so a wall stays a wall and a window stays a window across dozens of frames.
On the closed-loop test, where the camera returns to its starting point after a full trajectory, Mirage achieves the best scores on both PSNR and SSIM. This is the hardest test for spatial memory, because every tiny error accumulates over the entire path.
The efficiency numbers are striking. On a single H100, Mirage reads from its cache at roughly 0.25 seconds per frame after the first chunk, and the cache grows by less than 0.5 MiB per chunk. RGB-cache baselines like Spatia need several seconds per frame and multiple gigabytes of memory for the same trajectory.
How It Works (Without the Math)
Here is the simplified pipeline:
- Seed the memory. Mirage encodes the starting image into the diffusion model’s latent space, estimates depth, and back-projects each latent feature into a 3D location. Each point in the cache stores a full 48-channel feature vector, not just three RGB values.
- Read from memory. For each new camera position, Mirage projects the 3D cache directly onto the target camera at latent resolution. No rendering to pixels. No re-encoding. A single projection operation.
- Generate new frames. The projected features are injected into the video diffusion model via a ControlNet-style side branch. The model generates the new chunk of frames using the memory as a geometric guide.
- Update the memory. After generation, Mirage estimates depth on the new frames, filters out moving objects and sky, re-encodes the static parts, and back-projects them back into the 3D cache. The memory grows chunk by chunk.
Moving objects are deliberately excluded from the persistent memory because their geometry is unreliable. A person walking through a room should not be frozen into the walls. This is the current limitation: scenes with pervasive motion benefit less from the cache than quiet interiors.
What to Watch For
Mirage is built on Alibaba’s open-source Wan2.2 video model and adapts it with a ControlNet branch and LoRA adapters. The code and weights are on Microsoft’s GitHub under an MIT license. That means you can run this yourself, modify it, and build on it without asking permission.
The obvious next step is persisting dynamic content. Right now, Mirage remembers where the furniture is but not where the people are. A future version that can track and store moving objects across chunks would unlock fully interactive world simulation, where characters persist in space alongside the environment.
The broader trajectory is clear. Video world models are converging with game engines. Google’s Genie 3 creates interactive environments in real time. Google’s Gemini Omni is positioned as a world model successor to Veo. Mirage’s latent memory approach makes spatially consistent video generation fast enough and memory-efficient enough to be practical, which is the step that was missing.
If you generate video walkthroughs, architectural fly-throughs, or any spatial content that needs to stay consistent across camera moves, this is the approach that makes it work. Not because it produces prettier frames (though it does), but because it remembers what it already generated and does not forget.
Want more like this? Join MOKU CLUB for free. Weekly resources, early access to new guides, and occasional templates you can actually use. Join below.
