Hi,
I have spent the last couple weeks testing out both transfer 2.5 and predict’s synthetic data generation for a simple pick and place policy with a humanoid.
I have been through all the documentation and suggested cookbooks and have tried a multitude of prompts, parameters and conditions and have been getting sub par results that are either not consistent with real physics and develop strange image artifacts or apply extremely basic augmentations akin to most editing softwares.
Transfer seems to acheive slightly more consistent results and so we have decided to continue forward with just transfer 2.5 but still only produces 70% of usable data. Training a policy with cosmos also did not seem to improve our generalisation or success rate by any noticable margin.
I am wondering what I am doing wrong and would appriciate some help with the following:
- Is the models not ready for use out of the box and require post training for any results?
- Does transfer only help in improving specific environments and not in generalisation?
- Are there certain prompts that help preserve realism?
- Can any negative prompts be implimented or does it have to be the same as the examples provided?
- Is there a way to balance control weights to ensure scene stability over anything else?
I know that a similar topic has been put forward but since there has been no current updates it would be great to hear back from the cosmos team.
Thanks alot.