Kai Zhang

@qywu Great Job! If we want to batch the input_ids, what should we pad? the 0 results are terrible.

In the paper, the author removed all bias. That's right

3.2. Residual Dense Block Wd,c is the weights of the c-th Conv layer, where the bias term is omitted for simplicity.

Thanks for the comprehensive and well-curated paper list! [LED: Lexicon-Enlightened Dense Retriever for Large-Scale Retrieval](https://arxiv.org/pdf/2208.13661.pdf) has been accepted to WWW'23.

@caiyinqiong Thanks!!

+1

Thank you so much for the great catch! You are 100% correct and I will fix these typos in our next version. Best, Kai

Hi, thanks for the question. Training stable diffusion with magicbrush leads sub-optimal models. Models can not successfully learn editing task with this amount of data.

Close for inactivity for now, feel free to re-open/reply if you have any questions.

Hi Benno, Thanks for pointing it out, I will have a look with the modified code on my side later. If it works on my side I will update it...