Hi, I haven’t been able to find any documentation on best methods for using multiple camera views of a scene for Cosmos 3 reasoner to interpret. I have experimented with creating a mosaic of the feeds, interleaving key frames for each feed at each time stamp, and unsuccessfully trying to pass multiple feeds in an api call resulting in error. Is there a specific way that the model is best with for doing so as there doesn’t seem to be any documentation on this like there is for the generation and predict models (where you can set up a schema for the feed locations), which I’m assuming is because the reasoner may just not be trained in a way that is good for this. Thank you.
Hi @Ryan Beaudoin — you’re right that Cosmos Reason is less documented for multi-view inputs than Generate/Predict. As of the current NIM release, the reasoner is primarily tuned for single-image or single-video reasoning, which is why the multi-camera schema from the other models doesn’t directly apply.
**What works today**
1. **Tile/mosaic with explicit camera labels.** Your first instinct was correct. Build a single frame that stitches the camera views together with visual separators and text overlays like `CAM_FRONT`, `CAM_LEFT`, etc. Cosmos Reason can then reason about spatial relationships between labeled regions. Keep the aspect ratio reasonable — extreme panoramas can degrade performance.
2. **Interleaved key frames per timestamp.** If you need temporal reasoning, send a short sequence (4–8 frames) as a video where each frame contains the same multi-camera mosaic layout. Consistent layout across frames helps the model track objects through time.
3. **Pre-prompt with a structured schema.** Tell the model exactly what to look for:
```
The image shows 4 camera views: front, rear, left, right.
For each detected object, report: camera, class, approximate distance, and motion direction.
```
**What doesn’t work yet**
Passing multiple separate image/video URLs in one API call. The reasoner endpoint doesn’t expose a native multi-camera tensor input the way the prediction model does. You’ll need to fuse feeds client-side before calling the API.
**Practical tip for automotive/robotics**
If you’re using Cosmos Reason for scene understanding in a loop, cache the mosaic layout and only update the pixel content. Stable formatting reduces token cost and improves consistency across frames.
Hope that helps — if you find a layout that works well for your scene, it’d be great to share the results back.