You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
While benchmarking our internal prompt weighting, I noticed that compel's pooled embeddings change with different prompt weighting.
I.e. pooled embeds of test != pooled embeds of (test)0.9. Compel tokenizes and encodes the prompt string, including weighting syntax, for the pooled embeds. This doesn't seem intuitive or expected.
I don't have the bandwidth to create a PR for this, but it should be quite simple to fix.
The text was updated successfully, but these errors were encountered:
@BitPhinix why do you think we need to pool the original prompt string? let's say i have a prompt "a girl with (green)0.2 hair" where i want "green" to be very slow, in the pooled prompt I want it to be small as well
While benchmarking our internal prompt weighting, I noticed that compel's pooled embeddings change with different prompt weighting.
I.e. pooled embeds of
test
!= pooled embeds of(test)0.9
. Compel tokenizes and encodes the prompt string, including weighting syntax, for the pooled embeds. This doesn't seem intuitive or expected.I don't have the bandwidth to create a PR for this, but it should be quite simple to fix.
The text was updated successfully, but these errors were encountered: