Practical prompt engineering & control tips for Real→Real augmentation (Transfer 2.5)

Omniverse Synthetic Data Generation (SDG)
1

Hi Cosmos team,

I’m evaluating Cosmos Transfer 2.5 for label‑preserving real‑to‑real augmentation
(weather, lighting, appearance changes while keeping scene geometry stable).

I’ve reviewed the official docs and Cookbook and run experiments using
segmentation / edge controls, guidance tuning, and image style references.

I’d appreciate clarification on a few practical usage questions:

  1. Prompting
    – Are there recommended prompt patterns for Real→Real augmentation that
    consistently avoid scene drift?
    – Are there prompt styles that look plausible but often cause failures?

  2. Negative prompts
    – The docs mention the model is tuned around certain negative prompts.
    Should users generally avoid modifying them, or can they be adjusted safely?

  3. Guidance vs control weights
    – Is there a practical guidance / multi‑control weight range that balances
    variation with structural stability?

  4. Image style reference
    – For weather/lighting changes, should most variation come from the style image
    rather than the text?
    – How important is camera/viewpoint similarity between input and reference?

Thanks a lot for any guidance or references.