Open Happ1ness-dev opened 1 week ago
I think this might be a bug/feature with the Advanced CLIP encoding nodes, since all my code does is call its functions. What BREAK does is it calls clip.tokenize() on each text chunk, then joins the chunks together to create a single set of tokens, and then passes it into the encoding function.
might be a bug/feature with the Advanced CLIP encoding nodes
Could be, but it happens only when I use BREAK, and it also seems to happen very randomly. In simple prompts like this, the last (second) BREAK seems to be affected, while in lengthy prompts, the first BREAK is affected. In both cases, I used only 2 BREAKs.
It's almost as if I'm crossing some token threshold 🤔 But that one is just a guess.
BREAK always pads the chunk to whatever the maximum for one chunk is (I think it was 75 or something), so you should hit this bug without breaks if your prompt exceeds that chunk size times whatever maximum number of chunks the code can handle.
Hello, I've found a pretty weird bug(?) when using
comfy++
inSTYLE()
. Weighting in some BREAKs seems to be getting ignored when used withcomfy++
clip encoding, for some reason.To Reproduce: Try something like this:
STYLE(comfy++) A dog BREAK (a cat:1) BREAK (a snake:9)
. The output should be all artifacts, but it's not. It doesn't change no matter what you put in place ofa snake
and no matter the weight.comfy
,A1111
andcompel
all work fine.Expected behavior: Process all tokens properly and not ignore the weighting after some BREAKs.