dvlab-research / Video-P2P

Video-P2P: Video Editing with Cross-attention Control
https://video-p2p.github.io/
355 stars 24 forks source link

The use of config #3

Closed Laidawang closed 1 year ago

Laidawang commented 1 year ago

hi, I am so amazed at your work,But I still have doubts about some parameters in config p2p.I want to know the specific role of blend_word and eq_params.

for EX.

thx

ShaoTengLiu commented 1 year ago

Hi,

  1. The blend_word determines the attention mask, which is described in our paper's Alg. 1 line 11.
  2. The eq_params is used here for re-weighting, as discussed in paper's section 4.2 (Attention re–weighting).
  3. For more details, you can also refer to prompt-to-prompt.
  4. In your first example, you can set the blend_word to None and eq_params to {'words': 'authentic', 'values': 5}. Please refer to this config.
  5. In your second example, there are two kinds of changes. One is word swap and another is prompt refinement (global). I think our code structure doesn't support it at this moment. However, you can try to modify some codes here to complete two changes sequentially.

Thanks.

Laidawang commented 1 year ago

Your explanation is so detailed, I really appreciate it.Thank you!