Clarification on shared vs. independent flow timesteps for video and action

[lhrrhl0419]

Hi Cosmos team,

Thanks for releasing Cosmos 3 and the training code. I have a question about the flow timestep sampling for joint video-action training.

The technical report states that noise levels are sampled independently for each modality. However, in the current code, when independent_action_schedule=False, the action sigma/timestep appears to reuse the vision sigma/timestep:

sigmas_for_action = sigmas if sigmas_action is None else sigmas_action

and the default configuration seems to set:

independent_action_schedule: false

Could you please clarify:

Were the released Cosmos 3 action/mid-training checkpoints trained with independent video/action timesteps (independent_action_schedule=True) or with shared timesteps?

Is the default False setting intentional for post-training only, or does it also reflect the original mid-training setup?

Thanks for the clarification!