stable-diffusion-webui-prompt-parser/scripts/prompt-refactor.py

# Created by Jake Carter @ https://github.com/JakeCarterDPM/stable-diffusion-webui-prompt-refactor
# Rewritten as Prompt Parser Analyzer based on advanced prompt parser
# Fixed emphasized handling to avoid 'Tree' object has no attribute 'type' error

import re
import gradio as gr
from modules import script_callbacks, scripts, shared
from modules.shared import opts
from collections import namedtuple
import lark
import random
from functools import lru_cache
import hashlib
from itertools import product
import os
import logging
import traceback

logger = logging.getLogger(__name__)

# Feature flags (can be overridden via env)
def _env_bool(name: str, default: str = "0") -> bool:
    v = str(os.getenv(name, default)).strip().lower()
    return v not in ("0", "", "false", "no", "off")

ALLOW_EMPTY_ALTERNATE = _env_bool("ALLOW_EMPTY_ALTERNATE", "0")
EXPAND_ALTERNATE_PER_STEP = _env_bool("EXPAND_ALTERNATE_PER_STEP", "1")
GROUP_COMBO_LIMIT = int(os.getenv("GROUP_COMBO_LIMIT", "100"))
CACHE_SIZE = int(os.getenv('PROMPT_PARSER_CACHE_SIZE', 4096))

# Lark Grammar
_alt_rule = r' "[" prompt ("|" prompt)* "]" ' if not ALLOW_EMPTY_ALTERNATE else r' "[" prompt ("|" [prompt])+ "]" '

_grammar = r"""
!start: (prompt | /[][():|]/+)*

prompt: (scheduled | emphasized | grouped
       | alternate | alternate_distinct
       | alternate1 | alternate2
       | top_level_sequence | sequence
       | compound | numbered | and_rule
       | plain | WHITESPACE)*

!emphasized: "(" prompt ")"
        | "(" prompt ":" prompt ")"
        | "(" prompt ":" NUMBER ")"
        | "[" prompt "]"

scheduled: "[" [prompt (":" prompt)+] "]" ":" NUMBER (step_range_list | reverse_flag | step_range_list reverse_flag)?
reverse_flag: "reverse" | "r"
step_range_list: step_range ("," step_range)*
step_range: NUMBER "-" NUMBER | NUMBER "%" "-" NUMBER "%"

alternate: """ + _alt_rule + r"""
!alternate_distinct: "[" prompt ("|" prompt)* "]!"
alternate1: (prompt) "|" (prompt)+
alternate2: (plain | compound) ("|" (plain | compound))+

grouped: "{" ((NUMBER_Q | prompt | sequence | grouped) ("," | "|")?)+ "}"

top_level_sequence: prompt ("::" sequence)+ "!!" ("," plain)?
sequence: prompt "::" prompt ("," | WHITESPACE)* nested_sequence* ("!" | ";")
nested_sequence: "::" prompt ("," | WHITESPACE)* ("!" | ";" | "~")

compound: /[a-zA-Z0-9]+(_[a-zA-Z0-9]+)+/
numbered: NUMBER_Q ("!" | "_")? (grouped | sequence | compound | and_rule | plain | alternate | alternate_distinct | alternate1 | alternate2)
and_rule: (plain | compound) ("&" (plain | compound))+
WHITESPACE: /\s+/
plain: /([^\\\[\]()&]|\\.)+/

%import common.SIGNED_NUMBER -> NUMBER
%import common.INT -> NUMBER_Q
"""

schedule_parser = lark.Lark(_grammar)

@lru_cache(maxsize=CACHE_SIZE)
def hash_tree(tree: lark.Tree | lark.Token) -> str:
    if isinstance(tree, lark.Tree):
        return hashlib.md5((tree.data + ''.join(hash_tree(c) for c in tree.children)).encode()).hexdigest()
    return hashlib.md5(str(tree).encode()).hexdigest()

def resolve_tree(tree: lark.Tree | lark.Token, keep_spacing: bool = True) -> str:
    if isinstance(tree, lark.Tree):
        children = []
        for child in tree.children:
            if isinstance(child, lark.Token) and child.type == "WHITESPACE":
                if keep_spacing:
                    children.append(" ")
                continue
            children.append(resolve_tree(child, keep_spacing))
        result = "".join(str(c) for c in children if c)
        return re.sub(r"[\s\u2028\u2029]+", " ", result).strip() if keep_spacing else result.strip()
    return str(tree).strip()

class ScheduleTransformer(lark.Transformer):
    def __init__(self, total_steps: int, current_step: int = 1, seed: int | None = 42):
        super().__init__()
        self.total_steps = total_steps
        self.current_step = current_step
        self.seed = seed
        self.rng = random.Random(seed) if seed is not None else random

    def start(self, args):
        return "".join(str(arg) for arg in args if arg)

    def prompt(self, args):
        return "".join(str(arg) for arg in args if arg)

    def plain(self, args):
        return args[0].value

    def compound(self, args):
        return "_".join(str(arg) for arg in args)

    def and_rule(self, args):
        return " and ".join(resolve_tree(arg, keep_spacing=True) for arg in args if resolve_tree(arg))

    def grouped(self, args):
        return ", ".join(resolve_tree(arg, keep_spacing=True) for arg in args if resolve_tree(arg).strip(" ,|"))

    def alternate(self, args):
        vals = []
        for arg in args:
            s = resolve_tree(arg, keep_spacing=True)
            if s or s == "":
                vals.append(s)
        return vals[(self.current_step - 1) % len(vals)] if vals else "empty_prompt"

    def alternate_distinct(self, args):
        options = [resolve_tree(arg, keep_spacing=True) for arg in args if resolve_tree(arg)]
        return self.rng.choice(options) if options else "empty_prompt"

    def alternate1(self, args):
        options = [resolve_tree(arg, keep_spacing=True) for arg in args if resolve_tree(arg)]
        return self.rng.choice(options) if options else "empty_prompt"

    def alternate2(self, args):
        options = [resolve_tree(arg, keep_spacing=True) for arg in args if resolve_tree(arg)]
        return self.rng.choice(options) if options else "empty_prompt"

    def numbered(self, args):
        quantity = int(args[0])

        distinct = False
        if len(args) > 1:
            mark = str(args[1])
            distinct = mark in ("!", "_")

        target = args[-1]

        options = []
        if isinstance(target, lark.Tree) and target.data in ("alternate", "alternate1", "alternate2"):
            for child in target.children:
                val = self.visit(child)
                if val:
                    options.append(val)
        elif isinstance(target, lark.Token):
            options = [resolve_tree(target, keep_spacing=True)]
        else:
            for child in target.children:
                val = self.visit(child)
                if val:
                    options.append(val)

        if not options:
            return "empty_prompt"

        if distinct:
            if quantity >= len(options):
                unique = self.rng.sample(options, len(options)) if options else []
                pad = self.rng.choices(options, k=quantity - len(unique)) if quantity > len(unique) else []
                selected = unique + pad
            else:
                selected = self.rng.sample(options, quantity)
        else:
            selected = self.rng.choices(options, k=quantity)

        return ", ".join(selected)

    def sequence(self, args, parent=None):
        owner = resolve_tree(args[0], keep_spacing=True) if parent is None else parent
        descriptors = [resolve_tree(arg, keep_spacing=True).strip(" ,~!;") for arg in args[1:] if resolve_tree(arg).strip(" ,~!;")]
        return f"{owner}: {', '.join(descriptors)}"

    def top_level_sequence(self, args):
        owner = resolve_tree(args[0], keep_spacing=True).strip()
        sequences = []
        trailing_text = []
        for child in args[1:]:
            if isinstance(child, lark.Tree) and child.data == "sequence":
                sequences.append(self.sequence(child.children, owner))
            elif isinstance(child, str) and child.strip() == "!!":
                continue
            else:
                t = resolve_tree(child, keep_spacing=True).strip(" ,")
                if t:
                    trailing_text.append(t)
        text = f"{owner} -> {', '.join(sequences)}"
        if trailing_text:
            text += f", {', '.join(trailing_text)}"
        return text

    def nested_sequence(self, args):
        elements = [resolve_tree(arg, keep_spacing=True).strip(" ,~!;") for arg in args[:-1] if resolve_tree(arg).strip(" ,~!;")]
        terminator = args[-1] if args and isinstance(args[-1], str) else None
        if terminator == "~":
            return self.rng.choice(elements) if elements else "empty_prompt"
        return f"[{' | '.join(elements)}]"

    def emphasized(self, args):
        prompt = args[0]
        if len(args) > 1:
            if isinstance(args[1], lark.Token) and args[1].type == "NUMBER":
                weight = float(args[1].value)
                return f"({prompt}:{weight})"
            else:
                second = args[1]
                return f"({prompt}:{second})"
        else:
            return f"({prompt}:1.1)"

    def scheduled(self, args):
        prompts = [arg for arg in args[:-1] if not isinstance(arg, lark.Token) or arg.type != "NUMBER"]
        number_node = args[-1]
        if isinstance(number_node, lark.Tree):
            number_node = resolve_tree(number_node, keep_spacing=True)
        try:
            weight = float(number_node)
        except ValueError:
            weight = 1.0

        boundary = int(weight * self.total_steps) if weight <= 1.0 else int(weight)
        boundary = max(1, min(boundary, self.total_steps))

        if not prompts:
            return "empty_prompt"
        if len(prompts) == 1:
            return f"({resolve_tree(prompts[0], keep_spacing=True)}:{weight})" if self.current_step >= boundary else ""
        step_increment = boundary / max(1, len(prompts))
        for i, prompt in enumerate(prompts):
            step = min(self.total_steps, int(i * step_increment)) if i < len(prompts) - 1 else self.total_steps
            if self.current_step <= step:
                return f"({resolve_tree(prompt, keep_spacing=True)}:{weight})"
        return f"({resolve_tree(prompts[-1], keep_spacing=True)}:{weight})"

class CollectSteps(lark.Visitor):
    def __init__(self, steps, prefix="", suffix="", depth=0, use_scheduling=True, seed=None):
        super().__init__()
        self.steps = steps
        self.prefix = prefix
        self.suffix = suffix
        self.depth = depth
        self.use_scheduling = use_scheduling
        self.seed = seed
        self.rng = random.Random(seed) if seed is not None else random
        self.schedules = []

    def visit(self, tree):
        if isinstance(tree, lark.Tree):
            method_name = f"visit_{tree.data}"
            method = getattr(self, method_name, self._default_visit)
            return method(tree)
        elif isinstance(tree, lark.Token):
            return self._visit_token(tree)
        return []

    def _default_visit(self, tree):
        schedules = []
        for i, child in enumerate(tree.children):
            if isinstance(child, lark.Token) and child.type == "WHITESPACE":
                continue
            pre = "".join(resolve_tree(c, keep_spacing=True) for j, c in enumerate(tree.children) if j < i and not (isinstance(c, lark.Token) and c.type == "WHITESPACE"))
            post = "".join(resolve_tree(c, keep_spacing=True) for j, c in enumerate(tree.children) if j > i and not (isinstance(c, lark.Token) and c.type == "WHITESPACE"))
            collector = CollectSteps(self.steps, prefix=self.prefix + pre, suffix=post + self.suffix, depth=self.depth + 1, use_scheduling=self.use_scheduling, seed=self.seed)
            child_schedules = collector.visit(child)
            schedules.extend(child_schedules)
        return schedules

    def _visit_token(self, token):
        if token.type == "WHITESPACE":
            return []
        return [[self.steps, self.prefix + str(token) + self.suffix]]

    def visit_plain(self, tree):
        text = resolve_tree(tree, keep_spacing=True)
        return [[self.steps, self.prefix + text + self.suffix]]

    def visit_top_level_sequence(self, tree):
        transformer = ScheduleTransformer(self.steps, 1, self.seed)
        text = transformer.transform(tree)
        return [[self.steps, self.prefix + text + self.suffix]]

    def visit_scheduled(self, tree):
        if not tree.children:
            return [[self.steps, self.prefix + "empty_prompt" + self.suffix]]

        prompts = [
            p for p in tree.children
            if not (isinstance(p, lark.Token) and p.type == "NUMBER")
            and not (isinstance(p, lark.Tree) and p.data in ("step_range_list", "reverse_flag"))
        ]
        number_node = next((p for p in tree.children if isinstance(p, lark.Token) and p.type == "NUMBER"), None)
        step_range_list = next((p for p in tree.children if isinstance(p, lark.Tree) and p.data == "step_range_list"), None)
        is_reverse = any(isinstance(p, lark.Tree) and p.data == "reverse_flag" for p in tree.children)

        weight = float(number_node.value) if number_node else 1.0

        def _clamp_step(x: int) -> int:
            return max(1, min(x, self.steps))

        step_intervals = []
        explicit_ranges = False
        if step_range_list:
            explicit_ranges = True
            for sr in step_range_list.children:
                if not (isinstance(sr, lark.Tree) and sr.data == "step_range"):
                    continue
                if len(sr.children) != 2:
                    continue
                start_txt = resolve_tree(sr.children[0], keep_spacing=False)
                end_txt = resolve_tree(sr.children[1], keep_spacing=False)

                def _to_steps(txt: str) -> int:
                    s = txt.strip()
                    if s.endswith("%"):
                        try:
                            return int(round(float(s[:-1]) / 100.0 * self.steps))
                        except ValueError:
                            return 1
                    try:
                        return int(round(float(s)))
                    except ValueError:
                        return 1

                start_step = _clamp_step(_to_steps(start_txt))
                end_step = _clamp_step(_to_steps(end_txt))
                if start_step < end_step:
                    step_intervals.append((start_step, end_step))
        else:
            num_prompts = len(prompts)
            boundary = _clamp_step(int(round(weight * self.steps)) if weight <= 1.0 else int(round(weight)))
            if num_prompts == 1:
                schedules = []
                schedules.append([boundary - 1, self.prefix + self.suffix])
                last_text = resolve_tree(prompts[0], keep_spacing=True)
                schedules.append([self.steps, self.prefix + last_text + self.suffix])
                return schedules
            if boundary < num_prompts:
                boundary = num_prompts
            step_size = boundary / num_prompts
            for i in range(num_prompts):
                start = _clamp_step(int(round(i * step_size)) + 1)
                end = _clamp_step(int(round((i + 1) * step_size)))
                if start < end:
                    step_intervals.append((start, end))

        if is_reverse:
            prompts = prompts[::-1]
            step_intervals = step_intervals[::-1]

        schedules = []

        if step_intervals and step_intervals[0][0] > 1:
            schedules.append([step_intervals[0][0] - 1, self.prefix + self.suffix])

        for i, (start, end) in enumerate(step_intervals):
            end = min(end, self.steps)
            if start < end:
                p = prompts[i]
                if isinstance(p, lark.Tree):
                    child_schedules = self.visit(p)
                else:
                    text = resolve_tree(p, keep_spacing=True)
                    child_schedules = [[self.steps, text]]

                for sched in child_schedules:
                    schedules.append([end, self.prefix + sched[1] + self.suffix])

        if step_intervals and step_intervals[-1][1] < self.steps:
            tail_text = resolve_tree(prompts[-1], keep_spacing=True)
            schedules.append([self.steps, self.prefix + tail_text + self.suffix])

        if not schedules:
            return [[self.steps, self.prefix + resolve_tree(tree, keep_spacing=True) + self.suffix]]

        return schedules

    def visit_alternate(self, tree):
        options = []
        for child in tree.children:
            if isinstance(child, lark.Token) and child.type == "WHITESPACE":
                continue
            child_schedules = self.visit(child)
            child_options = [
                sched[1].strip(" ,|")
                for sched in child_schedules
                if sched[1].strip(" ,|")
            ]
            options.append(
                child_options
                or [resolve_tree(child, keep_spacing=True).strip(" ,|")]
            )

        if not options:
            return [[self.steps, self.prefix + "empty_prompt" + self.suffix]]

        if EXPAND_ALTERNATE_PER_STEP:
            schedules = []
            for step in range(1, self.steps + 1):
                option = options[(step - 1) % len(options)]
                for sched in option:
                    schedules.append([step, self.prefix + sched + self.suffix])
            return schedules
        else:
            group = options[self.rng.randrange(len(options))]
            choice = self.rng.choice(group) if group else "empty_prompt"
            return [[self.steps, self.prefix + choice + self.suffix]]

    def visit_alternate_distinct(self, tree):
        options = []
        for child in tree.children:
            if isinstance(child, lark.Token) and child.type == "WHITESPACE":
                continue
            child_schedules = self.visit(child)
            child_options = [
                sched[1].strip(" ,|")
                for sched in child_schedules
                if sched[1].strip(" ,|")
            ]
            options.append(
                child_options
                or [resolve_tree(child, keep_spacing=True).strip(" ,|")]
            )

        flat = [opt for group in options for opt in group]
        if not flat:
            return [[self.steps, self.prefix + "empty_prompt" + self.suffix]]

        selected = self.rng.choice(flat)
        return [[self.steps, self.prefix + selected + self.suffix]]

    def visit_alternate1(self, tree):
        return self.visit_alternate_distinct(tree)

    def visit_alternate2(self, tree):
        options = [resolve_tree(c).strip() for c in tree.children]
        combined_options = []
        for option in options:
            if "_" in option:
                combined_options.append(option)
            else:
                suffix = options[0].split("_")[-1] if "_" in options[0] else ""
                combined_options.append(f"{option}_{suffix}" if suffix else option)
        return [[self.steps, self.prefix + "|".join(combined_options) + self.suffix]]

    def visit_grouped(self, tree):
        all_options = []
        for child in tree.children:
            if isinstance(child, lark.Token) and child.type == "WHITESPACE":
                continue
            child_schedules = self.visit(child)
            child_options = [sched[1].strip(" ,|") for sched in child_schedules if sched[1].strip(" ,|")]
            all_options.append(child_options or [resolve_tree(child, keep_spacing=True).strip(" ,|")])
        out = []
        for i, combo in enumerate(product(*all_options)):
            if i >= GROUP_COMBO_LIMIT:
                break
            text = ", ".join(combo).strip()
            if text:
                out.append([self.steps, self.prefix + text + self.suffix])

        return out or [[self.steps, self.prefix + "empty_prompt" + self.suffix]]

    def visit_sequence(self, tree):
        transformer = ScheduleTransformer(self.steps, 1, self.seed)
        text = transformer.transform(tree)
        return [[self.steps, self.prefix + text + self.suffix]]

    def visit_nested_sequence(self, tree):
        elements = [resolve_tree(child, keep_spacing=True).strip(" ,~!;") for child in tree.children[:-1] if resolve_tree(child).strip(" ,~!;")]
        terminator = tree.children[-1].value if tree.children and isinstance(tree.children[-1], lark.Token) else None
        if terminator == "~":
            text = self.rng.choice(elements) if elements else "empty_prompt"
        else:
            text = f"[{' | '.join(elements)}]"
        return [[self.steps, self.prefix + text + self.suffix]]

    def visit_numbered(self, tree):
        quantity = int(tree.children[0])

        distinct = False
        if len(tree.children) > 1:
            mark = str(tree.children[1])
            distinct = mark in ("!", "_")

        target = tree.children[-1]

        child_schedules = self.visit(target)
        options = [
            sched[1].strip(" ,|")
            for sched in child_schedules
            if sched[1].strip(" ,|")
        ]
        if not options:
            options = [resolve_tree(target, keep_spacing=True).strip(" ,|")]

        if not options:
            return [[self.steps, self.prefix + "empty_prompt" + self.suffix]]

        if distinct:
            if quantity >= len(options):
                unique = self.rng.sample(options, len(options)) if options else []
                pad = self.rng.choices(options, k=quantity - len(unique)) if quantity > len(unique) else []
                selected = unique + pad
            else:
                selected = self.rng.sample(options, quantity)
        else:
            selected = self.rng.choices(options, k=quantity)

        return [[self.steps, self.prefix + ", ".join(selected) + self.suffix]]

    def visit_and_rule(self, tree):
        text = " and ".join(resolve_tree(c, keep_spacing=True) for c in tree.children if resolve_tree(c, keep_spacing=True))
        return [[self.steps, self.prefix + text + self.suffix]]

    def visit_emphasized(self, tree):
        prompt = resolve_tree(tree.children[0], keep_spacing=True)
        if len(tree.children) > 1:
            if isinstance(tree.children[1], lark.Token) and tree.children[1].type == "NUMBER":
                weight = float(tree.children[1].value)
                text = f"({prompt}:{weight})"
            else:
                second = resolve_tree(tree.children[1], keep_spacing=True)
                text = f"({prompt}:{second})"
        else:
            text = f"({prompt}:1.1)"
        return [[self.steps, self.prefix + text + self.suffix]]

    def __call__(self, tree):
        self.schedules = self.visit(tree)
        return self.schedules or [[self.steps, self.prefix + resolve_tree(tree, keep_spacing=True) + self.suffix]]


def at_step_from_schedule(step, schedule):
    if not schedule:
        return ""
    for end_step, text in schedule:
        if step <= int(end_step):
            return text
    return schedule[-1][1]

def at_step(step: int, prompt_or_schedule, *, steps: int | None = None,
            seed: int | None = 42, use_visitor: bool = True) -> str:
    if isinstance(prompt_or_schedule, list) and prompt_or_schedule and isinstance(prompt_or_schedule[0], list):
        return at_step_from_schedule(step, prompt_or_schedule)

    prompt = str(prompt_or_schedule)
    if steps is None:
        raise ValueError("steps is required when passing a prompt string to at_step(...)")
    sched = get_schedule(prompt, steps, True, seed, use_visitor)
    return at_step_from_schedule(step, sched)

@lru_cache(maxsize=CACHE_SIZE)
def get_schedule(prompt: str, steps: int, use_scheduling: bool, seed: int | None, use_visitor: bool = True):
    try:
        tree = schedule_parser.parse(prompt)
    except lark.exceptions.LarkError as e:
        logger.warning("Prompt parse failed: '%s' — %s", prompt, e)
        return [[steps, prompt]]

    collector = CollectSteps(steps, use_scheduling=use_scheduling, seed=seed)
    schedules = collector(tree)

    if not use_visitor:
        rebuilt = []
        for end, _ in schedules:
            transformer = ScheduleTransformer(steps, end, seed)
            text = transformer.transform(tree)
            rebuilt.append([end, text])
        return rebuilt

    return schedules

def analyze_prompt(input_prompt, steps, seed, use_visitor, output_mode, specific_step):
    tree_str = ""
    schedule_str = ""
    result = ""
    error_msg = ""
    try:
        tree = schedule_parser.parse(input_prompt)
        tree_str = tree.pretty()
    except lark.exceptions.LarkError as e:
        error_msg = f"Parse Error: {str(e)}\n\nPossible causes:\n- Unbalanced brackets or invalid syntax.\n- Missing colons, commas, or incorrect operators.\n- Suggestion: Validate grammar rules for scheduled, alternate, grouped, sequences, etc.\n- Ensure no invalid characters in plain text.\nTraceback: {traceback.format_exc()}"
        return result, tree_str, schedule_str, error_msg

    try:
        schedule = get_schedule(input_prompt, steps, True, seed, use_visitor)
        schedule_str = "\n".join([f"Up to step {end}: {text}" for end, text in schedule])
    except Exception as e:
        error_msg = f"Schedule Error: {str(e)}\nTraceback: {traceback.format_exc()}"
        return result, tree_str, schedule_str, error_msg

    if output_mode == "Full Schedule":
        result = schedule_str
    elif output_mode == "Parse Tree":
        result = tree_str
    elif output_mode == "At Specific Step":
        result = at_step_from_schedule(specific_step, schedule)
    return result, tree_str, schedule_str, error_msg

# Helper functions for converting prompt to specific formats
def convert_to_grouped(input_prompt):
    return f"{{{input_prompt}}}"

def convert_to_sequence(input_prompt):
    parts = [p.strip() for p in input_prompt.split(',')]
    if len(parts) < 2:
        return input_prompt + "::descriptor!"
    owner = parts[0]
    descriptors = ', '.join(parts[1:])
    return f"{owner}::{descriptors}!"

def convert_to_numbered(input_prompt, num, distinct):
    mark = "!" if distinct else ""
    return f"{num}{mark} {{{input_prompt}}}"

class PromptParserAnalyzerScript(scripts.Script):
    def title(self):
        return "Prompt Parser Analyzer"

    def show(self, is_txt2img):
        return scripts.AlwaysVisible

    def ui(self, is_txt2img):
        with gr.Group():
            with gr.Accordion("Prompt Parser Analyzer", open=False):
                with gr.Tab("Analyze Prompt"):
                    input_prompt = gr.Textbox(label="Input Prompt", lines=5)
                    steps = gr.Number(label="Steps", value=100, precision=0)
                    seed = gr.Number(label="Seed (for random elements)", value=42, precision=0)
                    use_visitor = gr.Checkbox(label="Use Visitor Mode (faster for simple prompts)", value=True)
                    output_mode = gr.Dropdown(choices=["Full Schedule", "Parse Tree", "At Specific Step"], label="Output Mode", value="Full Schedule")
                    specific_step = gr.Number(label="Specific Step (for At Specific Step mode)", value=50, precision=0)
                    analyze_btn = gr.Button("Analyze Prompt")
                    output_result = gr.Textbox(label="Analysis Result", lines=10)
                    output_tree = gr.Textbox(label="Parse Tree (Debug)", lines=5)
                    output_schedule = gr.Textbox(label="Full Schedule (Debug)", lines=5)
                    error_output = gr.Textbox(label="Errors & Suggestions", lines=5)

                    analyze_btn.click(
                        fn=analyze_prompt,
                        inputs=[input_prompt, steps, seed, use_visitor, output_mode, specific_step],
                        outputs=[output_result, output_tree, output_schedule, error_output]
                    )

                with gr.Tab("Format Converters"):
                    gr.HTML(value="<p>These helpers convert plain prompts to advanced formats. Read the explanations for usage.</p>")

                    with gr.Accordion("Grouped {} - For combinations of attributes"):
                        gr.HTML(value="<p>Wraps items in {}, generates combinations if | (alternates) inside. <br>Example: {red|blue, car|bike} resolves to 'red, car', 'red, bike', 'blue, car', 'blue, bike'. <br>Use for multiple independent attributes like colors, objects. Limit combos with GROUP_COMBO_LIMIT env var.</p>")
                        conv_group_input = gr.Textbox(label="Input for Grouped (e.g., red|blue, car|bike)", lines=2)
                        conv_group_output = gr.Textbox(label="Converted Grouped Prompt", lines=2)
                        conv_group_btn = gr.Button("Convert to Grouped")
                        conv_group_btn.click(convert_to_grouped, inputs=conv_group_input, outputs=conv_group_output)

                    with gr.Accordion("Sequence :: - For structured descriptions"):
                        gr.HTML(value="<p>Structures as owner::descriptors!. Nested with :: and ~ for random or ! for close. <br>Example: character::hair:ponytail, eyes:blue! -> 'character: hair ponytail, eyes blue'. <br>Use for hierarchical entities like characters, outfits. Close with ! or ;. Top-level with !!! for groups.</p>")
                        conv_seq_input = gr.Textbox(label="Input for Sequence (e.g., character, hair:ponytail, eyes:blue)", lines=2)
                        conv_seq_output = gr.Textbox(label="Converted Sequence Prompt", lines=2)
                        conv_seq_btn = gr.Button("Convert to Sequence")
                        conv_seq_btn.click(convert_to_sequence, inputs=conv_seq_input, outputs=conv_seq_output)

                    with gr.Accordion("Numbered N! or N_ - For repeats"):
                        gr.HTML(value="<p>N{group} repeats N times, ! or _ for distinct (no repeats). <br>Example: 3! {a|b|c} -> 'a, b, c' (unique sample). <br>Use for multiples like '3 cats'. If N > options, pads with repeats unless distinct.</p>")
                        conv_num_input = gr.Textbox(label="Input for Numbered (e.g., a|b|c)", lines=2)
                        conv_num_num = gr.Number(label="Number (N)", value=3, precision=0)
                        conv_num_distinct = gr.Checkbox(label="Distinct (! or _)", value=True)
                        conv_num_output = gr.Textbox(label="Converted Numbered Prompt", lines=2)
                        conv_num_btn = gr.Button("Convert to Numbered")
                        conv_num_btn.click(convert_to_numbered, inputs=[conv_num_input, conv_num_num, conv_num_distinct], outputs=conv_num_output)

        return [input_prompt, steps, seed, use_visitor, output_mode, specific_step, analyze_btn, output_result]

def on_ui_settings():
    section = ('prompt-parser-analyzer', "Prompt Parser Analyzer")

script_callbacks.on_ui_settings(on_ui_settings)