denoising-diffusion-pytorch
denoising-diffusion-pytorch copied to clipboard
Implementation of Denoising Diffusion Probabilistic Model in Pytorch
Hi Phil, It seems that the classifier guidance is missing for DDIM sampling. The `cond_fn` is not used in the `ddim_sample` function. https://github.com/lucidrains/denoising-diffusion-pytorch/blob/0dee2d8eebe251e666ef60c1b48d3333a18f4065/denoising_diffusion_pytorch/guided_diffusion.py#L666
Hi Phil, In your classifier free guidance file, you use `self.null_classes_emb = nn.Parameter(torch.randn(dim))`, which is random control information. In `forward_with_cond_scale` and `forward` ``` null_logits = self.forward(*args, cond_drop_prob = 1., **kwargs)...
May I ask why in [classifier_free_guidance.py](https://github.com/lucidrains/denoising-diffusion-pytorch/blob/main/denoising_diffusion_pytorch/classifier_free_guidance.py) you let `null_classes_emb=nn.Parameter(torch.randn(dim))` instead of `torch.zeros(dim)`, because the latter one seems more reasonable.
Hello, I noticed that data and all_images were used when calculating fid. However, these two images do not seem to be paired. Although all_images looks normal, fid is always around...
When the model iterated about 26,000 times, I receieve the following warning: [W pthreadpool-cpp.cc:90] Warning: Leaking Caffe2 thread-pool after fork. (function pthreadpool) All other configurations are defualt as the given...
will it automaticly load from the last checkpoint?
Hi, thanks so much for the very nice implementation of the diffusion model. I am wondering whether the repo contains the function of computing likelihood of the samples? Thanks
Hi, thanks for providing this code. I'm trying to reproduce the CIFAR10 results from the original DDPM paper. I use 3x32x32 images, all the CIFAR data (50k frames), 2000 epochs...
With the current implementation, is it possible to simply denoise a blurry image and not generate an image from raw white noise ? If it is possible, how can I...
Hi! Great works!! Really learn a lot!!! Thank you so much!!!!!! Just wondering, will the repo be able to do the img2img trick? Or any hint or snippet of code...