Wenqiang Sun

Oh, I am just thinking about this issue. The elevations and azimuths of training data may influence the inference performance. For example, in this work, they use zero123++ to generate...

I think that you can add a dictionary in the class AttentionStore to collect the cross-attention map for each step and add the cross-attention map of each step to the...

Oh, sorry to reply so late. Thanks for your suggestion! I have solved this problem and achieved high-quality generation results. I also find that the coefficient of LPIPS loss is...

@junwuzhang19 Hi, sorry to reply so late. I think you can adjust the lr to a smaller value. I use 1e-5 to 5e-5, and the training process is very stable....

@junwuzhang19 yes, I also find that 30 epochs will only take about one day using 8 A100 GPUs. Based on my results, I think that if you want to achieve...

@chenguolin Oh, that's amazing! Actually, I only train the LGM using the big version config. I am wondering whether you have compared the final performance on the test dataset using...