search

ManiCM-fast / ManiCM

ManiCM: Real-time 3D Diffusion Policy via Consistency Model for Robotic Manipulation
https://manicm-fast.github.io
MIT License
62 stars 5 forks source link

num_inference_step #5

Open kangkang2626 opened 1 month ago

kangkang2626 commented 1 month ago

hi, great job, have you try to setting the num_inference_step=1 in original DP3? The num_inference_step is set to 10 in this experiment, but i found that num_inference_step=1 is also work well. If setting the num_inference_step=1 in DP3, the inference time is almost same.

runjie-yan commented 1 month ago

hi, great job, I have samilar concerns on the inference steps. I also found that DP3 seems to work well with num_inference_step=1. I think one possible reason is the generated datasets in the virtual benchmark is not diverse enough and maybe training diffusion model on it is even unnecessary?

gao-zifeng commented 3 weeks ago

We did try setting fewer inference steps on dp3, with num_inference_step=4 resulting in an average success rate of 72.4%.