Dewei Zhou

@AZZMM Thank you for your interest in our work. We are recently expanding MIGC to MIGC++ with more comprehensive functions. We expect to submit a report to arxiv in May...

pre_attn = fuser_info['pre_attn'] # (B*PN, heads, HW, 77) BPN, heads, HW, _ = pre_attn.shape pre_attn = torch.sum(pre_attn[:, :, :, 1:], dim=-1) # (B*PN, heads, HW) H = W = int(math.sqrt(HW))...

@AZZMM In the training, we don't need negative prompts. BPN means Batch * Phase_num, the Phase_num contain {global prompt, instance1_desc, instance2_desc, ..., instanceN_desc}. We use the first two 16*16 attn-maps...

@AZZMM vanilla cross attention

@WUyinwei-hah We have already completed the writing of the MIGC++ paper, and we will submit it in the next few days. Then, we will proceed to consider the open-source work...

"cat" and "dog" are two very similar tokens, which can easily lead to attribute leakage during cross-attention. You can increase NaiveFuserSteps to 50 (i.e., consistent with num_inference_steps=50) to avoid attribute...

@yuntaodu Thank you for your interest in our work. As a result, have you set NaiveFuserSteps to be consistent with num_inference_steps to avoid attribute leakage to the greatest extent?

Thank you for your interest! Our training code was implemented under Huawei's internal training framework and would require a considerable amount of time to organize. Therefore, in the short time,...

Thank you very much for recognizing our work.😄 In fact, I have already been training the SDXL version of MIGC. However, the SDXL model has a larger number of parameters...

@fritol You can try changing the base model to get better generation quality.