chq1155/a-survey-on-generative-diffusion-model

a-survey-on-generative-diffusion-model's Issues

Add new paper "DeTiME: Diffusion-Enhanced Topic Modeling using Encoder-decoder based LLM"

Hi,

Thanks for your wonderful survey!
Would you mind adding our paper published in EMNLP 2023 as findings in section 4.2.1 Natural Language Processing? It leverages diffusion for topic guided text generation.
Title: DeTiME: Diffusion-Enhanced Topic Modeling using Encoder-decoder based LLM
Paper link: https://arxiv.org/abs/2310.15296

Add new reference

Please consider adding our paper "Diffusion Probabilistic Model Made Slim" arxiv, which has just been accept by CVPR2023.

Best.

Misunderstanding contents from the survey paper

Thank you very much for your review paper, this article has benefited me a lot. These references are also invaluable. In the process of reading through the article, I encountered some questions and hope to get your answers. The paper version I read is from arxiv.

1. at Eq. $(3)$, I think the formula of the paper has one more item $F_{0t}$, and it should be $F(x,\sigma)=F_{s1}(x_s, \sigma_{s1})\circ F_{ts}(x_t,\sigma_{ts})\circ F_{0t}(x_0,\sigma_{0t})$. Is that right?
2. For the annealed Langevin dynamics algorithm, there are a duplicated for loop $\text{for } i =1 \dots L \text{ do}$, is the second one should be replaced with $\text{for } t =1 \dots T \text{ do}$? And the random noise notation is also confusing.
3. Maybe the reference [94] is not right for CCDF but Cold Diffusion, which is Chung H, Sim B, Ye J C. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022: 12413-12422.

Looking forward to your reply. Thank you very much!

Best.

delete

Could you add our CVPR 2023 paper about video generation?

Hi, @chq1155, thanks a lot for your efforts to collect these works! I wonder whether you can add our CVPR 2023 paper "Conditional Image-to-Video Generation with Latent Flow Diffusion Models", which applies Diffusion models to synthesize optical flow for video generation. The links of our paper and code are paper_link and code_link.