comfyanonymous
Broken prompt weighting
Expected Behavior
prompts person (smiling:0.0) & person should be similar
In other worlds:
prompt = person + smiling * 0.0 => prompt = person
These are outputs from InvokeAI
person smiling
person
person (smiling)0.0
person (smiling)1.0
Actual Behavior
Steps to Reproduce
Debug Logs
Nothing important here.
Other
No response
It's also important to note that person (smiling:0.0) also outputs a person smiling.
It's also important to note that
person (smiling:0.0)also outputs a person smiling.
What about the space between person and smiling? As far as I know, there's a difference between person and person , which means they will tokenize differently. You should put the space in the parenthesis with your weighted word/phrase. That way, if they are equal to 0.0, it should remove them, along with the empty space.
I see your point and I will bring these examples. But I suspect that “person ” should be closer to “person” than “person smiling”.
Em qui., 22 de ago. de 2024 às 21:08, RandomGitUser321 < @.***> escreveu:
It's also important to note that person (smiling:0.0) also outputs a person smiling.
What about the space between person and smiling? As far as I know, there's a difference between person and person , which means they will tokenize differently. You should put the space in the parenthesis with your weighted word.
— Reply to this email directly, view it on GitHub https://github.com/comfyanonymous/ComfyUI/issues/4550#issuecomment-2305944331, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAIFAVAOAZDUBP5QJLZ37KTZSZ4QHAVCNFSM6AAAAABM65IGJ6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDGMBVHE2DIMZTGE . You are receiving this because you authored the thread.Message ID: @.***>
Interpretation of weighting is not that simple. Each tool has its own interpretation of weighting.
I recommend reading the Intuition behind weight interpretation methods part in this repo.
https://github.com/BlenderNeko/ComfyUI_ADV_CLIP_emb
