Skip to main content
Use Cosmos3 Nano Sim Transfer to explore visual variation without recording a new robot trajectory. For example, render the same tabletop motion with bright overhead lighting, dim side lighting, or a different tabletop material. Keep your control sequence fixed and change the scene description.

​
Run a comparison

Use the cookbook appearance-variant workflow to render the same prepared depth with different captions and separate output directories. For complete frame exports, use its incremental PNG example. Neither example provides durable batch jobs or dataset management.

​
Compare variants

  1. Render a depth sequence with a caption describing the scene and lighting.
  2. Wait for generation to finish. Reuse the uploaded source and seed, changing only the caption to compare appearance conditions.
  3. Inspect whether task-relevant objects and geometry remain consistent with the simulator.
  4. Evaluate whether those observations help your training or evaluation task before expanding the dataset.
For this workflow, sustained throughput and usable output matter more than interactive latency. The uploaded-video workflow streams a preview. To collect complete frame batches, use the step interface and save each RGB PNG result alongside its simulator timestamps, actions, camera settings, and scene description. Start a new episode for each variant. The two rendering modes do not promise identical pixels for the same seed.

​
Keep labels aligned

Generated appearance can change task-relevant details. Check object identity, boundaries, and visibility before reusing simulator labels. Retain the original depth and state so you can audit samples. Dataset storage, job scheduling, and training remain in your application.