search

xuekt98 / BBDM

BBDM: Image-to-image Translation with Brownian Bridge Diffusion Models
MIT License
284 stars 30 forks source link

DDIM support #25

Open eslambakr opened 1 year ago

eslambakr commented 1 year ago

Hello @xuekt98

Thanks for sharing the source code for your amazing work!

I have a quick question regarding the DDIM part. In Section 3.2 in the paper, you have mentioned that DDIM could be supported in your framework; BBDM. Also, in Table 4, you have tried different sampling steps using the DDIM. However, I can not see the part of the code that supports DDIM. Thus, it would be appreciated if you could point it out to me please.

Your answer will be much appreciated! Thanks in advance!

xuekt98 commented 1 year ago

Please refer to the sample function in BrownianBridgeModel.py

eslambakr commented 1 year ago

Thanks @xuekt98 for the prompt response!

Yes, I see but what I mean is that there is no special implementation for the DDIM in your code. As the sampling function is the same regardless the number of steps. So, is my understanding is correct, i.e., no difference in the sampling between DDPM and DDIM in your architecture)?

Also in Table 4, the reported results for the same model trained using 1K steps and the only difference is the sampling steps only?

xuekt98 commented 1 year ago

actually, it is easy to notice that ddpm sampling can be seen as a special case of ddim sampling. and the step in training is fixed to 1k and we only change different steps of sampling

---- Replied Message ---- | From | @.> | | Date | 10/18/2023 14:17 | | To | @.> | | Cc | @.>@.> | | Subject | Re: [xuekt98/BBDM] DDIM support (Issue #25) |

Thanks @xuekt98 for the prompt response!

Yes, I see but what I mean is that there is no special implementation for the DDIM in your code. As the sampling function is the same regardless the number of steps. So, is my understanding is correct, i.e., no difference in the sampling between DDPM and DDIM in your architecture)?

Also in Table 4, the reported results for the same model trained using 1K steps and the only difference is the sampling steps only?

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: @.***>