Hi, thank you for your amazing project.
I have a question about the diffusion steps in DiT module. I noticed that InterVLA-M1 levearge DDIM as the policy, unlike the flow-matching based method in GR00T. My questions are:
- How many steps does M1 need for efficient inference?
- Is DDIM much better than flow-matching based methods?
Hi, thank you for your amazing project.
I have a question about the diffusion steps in DiT module. I noticed that InterVLA-M1 levearge DDIM as the policy, unlike the flow-matching based method in GR00T. My questions are: