search

Rongjiehuang / FastDiff

PyTorch Implementation of FastDiff (IJCAI'22)
408 stars 64 forks source link

Pre-defined noise schedule #9

Closed yahshibu closed 2 years ago

yahshibu commented 2 years ago

Thank you for making your code publicly available!

I have a question about the pre-defined noise schedule. The paper says β = Linear(1e−4, 0.005, 1000) was used (Table 7). On the other hand, the default setting of this repository is β = Linear(0.000001, 0.01, 1000). https://github.com/Rongjiehuang/FastDiff/blob/f737e17e344923a7ec60507b289abd756f76000a/modules/FastDiff/config/FastDiff.yaml#L52-L54

Which setup is your recommendation?

To my understanding, the noise-predictor-derived noise schedule depends on a trained score network. It means the noise-predictor-derived noise schedule depends on the pre-defined noise schedule. Is this correct?

Rongjiehuang commented 2 years ago

Hi, this schedule is not such sensitive. We have tested bothβ = Linear(1e-6, 0.005, 1000) (in paper), or β = Linear(1e-6, 0.01, 1000) (this repo).

yahshibu commented 2 years ago

Hi, thank you for answering my question! Got it. It is nice that the schedule is not so sensitive!