Newbeeer
commented
1 year ago Hi,
It seems that you are using PFGM in research. We are excited to inform you that we release the PFGM++ paper ( https://arxiv.org/abs/2302.04265 ) and code (https://github.com/Newbeeer/pfgmpp) today.
😇 Improvements over PFGM / Diffusion Models:
- No longer require the large batch training target in PFGM, thus enable flexible conditional generation and more efficient training!
- More general D∈R+ dimensional augmented variable. PFGM++ subsumes PFGM and Diffusion Models: PFGM correspond to D=1 and Diffusion Models correspond to D→∞.
- Existence of sweet spot D* in the middle of (1,∞)!
- Smaller D more robust than Diffusion Models ( D→∞​)
- Enable the adjustment for model robustness and rigidity!
- Enable direct transfer of well-tuned hyperparameters from any existing Diffusion Models ( D→∞ )​
​We recommend using the new PFGM++ framework for your research :). According to our internal application of PFGM++ to other domains, such as graph generation and RNA sequence, PFGM++ works much better than its special case PFGM, and does not require careful hyperparameter tuning!
Hope you enjoy PFGM++ :)
Accident...