In the "PerceiverAttention block", the key and value have only one token?

instantX-research / InstantID

InstantID: Zero-shot Identity-Preserving Generation in Seconds 🔥

https://instantid.github.io/

Apache License 2.0

11.16k stars 813 forks source link

Open shoutOutYangJie opened 6 months ago

shoutOutYangJie commented 6 months ago

face embeding token as key and value, however, its length is 1.

haofanwang commented 6 months ago

It doesn't matter. The Resampler module will project the original face embedding into specified dimension and num_tokens.