gongyeliu / stylecrafter Goto Github PK

Dear author,

I would like to know in detail about the BLIP2 caption preprocessing method used in the paper.

1. Which BLIP2 checkpoint was used? Was there no difference in performance between the checkpoints?
2. What regular expressions were used?
3. Whether the captions were processed by BLIP-2 itself or Python source code?

Your paper inspires me. Thank you, in advance.

Thanks for your great work!
I would like to know what format the training data is in.

Can you provide the CLIP score evaluation code?

Thank you.

does this use https://github.com/google/style-aligned?

1. 请问大佬，我如果只想得到文本+参考的风格化图像得到风格化的图像模型，有好的项目可以参考进行训练吗，因为我发现我们项目代码基础模型加载的是文本生成视频的模型videocrafter_t2v_320_512，那我只想生成风格化图片，没有必要使用视频生成模型？
   
2. 我发现这里风格化和IP-Adapter（https://github.com/tencent-ailab/IP-Adapter）不同，项目中风格化更多是提取参考图的风格+ 文本提升内容，IP-Adapter更多倾向于参考图作为底图进行绘制

Hi, thanks to share free.

When i am looking to your work, wonder to To-Do list.
How to remove watermark?

Thank you.

Can you train detail "Style Embedding Extraction"?

1. Especially, a trainable "Q-Former", how to train this?

2. Can you publish train code?

Thank you.