a workflow which includes my custom node (https://github.com/levzzz5154/ComfyUI-MOSS-PromptWeight/) for MOSS image prompt weighting
use the default syntax -> (string of tokens:weight)
a little caveat is that you don't always need to backslash-escape parentheses if there's no weight, so e.g. you write oil painting (medium) instead of oil painting \(medium\)
if log is too weak for your taste change it to linear
it's also possible to prompt weight bboxes so the model doesn't follow them super strictly, i'll implement an easier way of doing it later
Description
FAQ
Comments (11)
Working perfectly for me, thank you :)
I enjoy this trend of slamming prompt weights back into models as soon as possible, thanks a lot
Ok, not sure if you take feedback or not, but it breaks past 3, and unlike Krea2 Prompt Weight (KJNodes) this node does not support (I think?) two tokens at once, see (big hips, slim waist:3), if you could fix these two it would be aamazing!
err yeah some bugs were found, it breaking at high values is kind of expected if you're using linear mapping; and two tokens should be supported... well, i figured out how to prompt weight bboxes, so it should work
however, do update your install
Getting very close to my Krea 2 experience here, but with the benefit of not needing these shitty refusal LoRA's either, plus no 'bimbo' LoRA's needed either, thanks man!
Working great with three tokens :)
ok within the last 14 minutes it has been sprayed away"based on the krea2-weights in kjnodes"
yet for some reason it features this gem:
decoder = codecs.getincrementaldecoder("utf-8")(errors="replace")
prompt += decoder.decode(bytes(hf.byte_decoder[c] for c in t))
which blows up with
AttributeError: Qwen2Tokernizer has no attribute byte_decoder.