Open Meerkov opened 12 months ago
I used the following tool to help debug https://www.lark-parser.org/ide/
According the prompt Lark script, this seems to be caused in part by the following:
[cat:-1]
-> this produces a schedule which begins using cat
at step -1. No clue why this is different than 0.
[cat:0]
-> this produces a schedule which begins using cat
at step 0. The weight is presumably 1.0.
{cat:0}
-> this produces plain text {cat:0}
with no special meaning.
(cat:0)
-> this produces an emphasis with cat
, but the weight is some non-zero amount due to a bug (reported elsewhere).
cat:0
-> this actually negates the cat
keyword correctly.
You want to produce an image without a cat ... and write cat into the prompt ... ? 🤔
You want to produce an image
@Woisek please don't comment unless you have something helpful to say. The bug has nothing to do with me "wanting to produce an image". It's about the bracket system being a mess and clearly not working as expected.
In short: prompt : X -> correctly weights the prompt with X weight (prompt:X) -> doesn't
Is there an existing issue for this?
What happened?
The following produce distinct images
[cat:0]
{cat:0}
-> Edit: this is actually 5 tokens(cat:0)
cat:0
These all are identified as a single token. The weight is the same in all cases. Only the last one produces a result which is recognizable as containing no cat, though the latent noise mid-generation shows a cat temporarily.
Steps to reproduce the problem
cat:-1
(cat:-1)
[cat:-1]
The same issue appears when using
-1
weights. Only by using no brackets can you get an image which properly has no cat.What should have happened?
It's clear that there is a bug in the system for weighting brackets, when compared to the same prompt with no brackets.
Sysinfo
.
What browsers do you use to access the UI ?
No response
Console logs
Additional information
No response