Open loboere opened 4 months ago
I would implement it but I don't understand what part of the code represents the id
I would implement it but I don't understand what part of the code represents the id
Thank you very much for your interest in our work. To understand the meaning of these tensors, you may need to set a breakpoint and refer to the article. The ID information is encoded in "ctx". If you want to mix the IDs, you should first obtain "ctx_id1" and "ctx_id2". Then, you can calculate "x" using the following formula:
x = x + (self.self_attn.self_attn_first(y, ctx_id1)) * self.share_cache['similarity'] + (1 - self.share_cache['similarity']) * self.self_attn.self_attn_first(y, ctx_id2)
if you encounter any difficulties, please don't hesitate to ask me for help.
I don't understand how to load two IDs, ctx was generated from y and y from x, I don't know what x is, could you give me some code to load two IDs?
Do you plan to publish the code for Identity Mixing?
Do you plan to publish the code for Identity Mixing?
Absolutely, this would be a modification to the existing code. I will add this function soon.
It's been a while, please add it as soon as possible
any news?
+1
please @jshilong add it as soon as possible
This code does not implement Identity Mixing. You can implement it by combining two ID results on https://github.com/ali-vilab/FlashFace/blob/a3afe3f5a6851104ae0895dfdb207cfbdc978ea8/flashface/all_finetune/models/reference_unet.py#L96 and https://github.com/ali-vilab/FlashFace/blob/a3afe3f5a6851104ae0895dfdb207cfbdc978ea8/flashface/all_finetune/models/reference_unet.py#L108