[Feature]: Merge positive and negative prompts into single prompt

[CzBuCHi]

Feature description

Currently it is impossible to have dynamic negative prompt based on positive prompt

My proposal is to use negative weights to describe stuff for negative prompt - system would then just extract those for negative prompt:

prompt: foo, (bar:-1)

would be equivalent to current:
positive: foo
negative: bar (or (bar:1) for simplicity)

Reasoning: wildcards

wildcard example: lets say i have wildcard file color.txt

red
green
blue

then currently im unable to specify 'if red is selected add purple to negative prompt'

if prompts got merged i could write my wildcard file like this:

red, (purple:-1)
green
blue

and system would recognized that purple belongs to negative prompt

result would be even more powerful wildcards ...

PS: it would be nice to show final prompt in log not just list of possible wildcards, but that just cosmetic ....

Version Platform Description

windows 11, latest version of sdnext

[vladmandic]

valid idea, but this would be quite involving with high risk of regressions.

[CzBuCHi]

i think it could be done before current code - sort of preprocessor:

basically instead of reading positive and negative prompts directly from UI code will read single prompt from UI and then call new method to extract positive and negative prompts from that ...

all the new magic would be in this separate method -> not affecting any part of current code ...

and for UI changes: it would basically be just removing negative prompt input ...

---

after chat with GPT i think this code would do the trick: (but im not python programmer, so i cannot verify code)

import re

# split into tokens
# "(foo, bar:3), baz" → ["(foo, bar:3)", "baz"]
def split_top_level(s: str):
    parts = []
    buf = []
    depth = 0

    for ch in s:
        if ch == ',' and depth == 0:
            parts.append(''.join(buf).strip())
            buf = []
            continue

        if ch == '(':
            depth += 1
        elif ch == ')':
            depth -= 1

        buf.append(ch)

    if buf:
        parts.append(''.join(buf).strip())

    return parts

# separate positive and negative tokens
#  "foo, (bar:0.5), (fizz:-1), (buzz:-0.5), (foo, bar:3)" → "foo, (bar:0.5), (foo, bar:3)", "(fizz:1), (buzz:0.5)"
def split_prompt(s: str):
    pattern = re.compile(r'^\((?P<term>.+?):(?P<weight>-?\d+(?:\.\d+)?)\)$')

    positive = []
    negative = []

    for token in split_top_level(s):
        m = pattern.match(token)

        if not m:
            positive.append(token)
            continue

        term = m.group('term')
        weight = float(m.group('weight'))

        if weight >= 0:
            positive.append(token)
        else:
            negative.append(f"({term}:{abs(weight)})")

    return ", ".join(positive), ", ".join(negative)

and theoretical usage:

-pos = read_positive_input()
-net = read_negative_input()

+prompt = read_positive_input()
+pos, neg = split_prompt(prompt)

image = generator(pos, neg)

ofc there may be errors (malformed prompt, etc.) - in that case i would simply do pos = prompt, neg = "" and print error to console ...

[vladmandic]

actually replacing positive/negative prompts from ui is not an option as 99.9% of users expect them to be separate.
which means this would need to be pre-processor for positive prompt that removes items from positive prompt and places them in negative prompt instead.
but then there are tons and tons of edge cases. what if model doesn't support negative prompts or is running in guidance scale <= 1 (which means negative disabled)? we would modifying positive prompt to place things in negative prompt would basically make items disapear. and thats just one example.
also, parser would have to be very very precise. "there may be errors and print to console" is not what most users would expect.
all in all - very big risk for very edge use case.