StyleRef-styleref-comfyui/scripts/build_workflows.py

#!/usr/bin/env python3
"""
Generate the shipped workflow templates in workflows/.

Hand-maintaining four ComfyUI workflow files means hand-maintaining node ids,
link ids, and slot indices across all of them — the kind of bookkeeping that is
silently wrong until someone drags the file onto a canvas. This builds them from
a description instead, so the ids are correct by construction.

Layout policy: node positions are never guessed. Every node's height is the
*measured* minimum for its type (see `MIN_H`), and nodes stack per column with a
fixed vertical gap — so no template can ship with overlapping nodes. A test
asserts that invariant on the committed files.

Instruction policy: user-facing instructions live in visible core
`Note` nodes on the canvas, and they stay **short** — a note is a signpost, not
a manual. Anything longer belongs in the README or docs.styleref.io.
`extra.styleref.notes` carries only a one-line provenance string — ComfyUI never
renders it, so it must not carry anything a user needs to read.

Run after changing a node's inputs or outputs:

    python scripts/build_workflows.py

CI checks the committed files match this script's output.
"""

from __future__ import annotations

import json
import os
import sys
from typing import Any

_PACK_ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
if _PACK_ROOT not in sys.path:
    sys.path.insert(0, _PACK_ROOT)

# Single-sourced from the node itself (the Facets output list is data-driven
# and long — duplicating it here would guarantee drift).
from styleref_nodes.facets import StyleRefFacets  # noqa: E402

WORKFLOW_DIR = os.path.join(_PACK_ROOT, "workflows")

PROVENANCE = "Generated by scripts/build_workflows.py — edit the script, not this file."

# Slot layouts, mirroring each node's RETURN_TYPES / INPUT_TYPES order.
OUTPUTS = {
    "StyleRefLoad": [
        ("style", "STYLEREF_STYLE"),
        ("search_results", "STRING"),
    ],
    "StyleRefApply": [("positive", "STRING"), ("negative", "STRING"), ("attribution", "STRING")],
    "StyleRefFacets": list(zip(StyleRefFacets.RETURN_NAMES, StyleRefFacets.RETURN_TYPES)),
    "StyleRefReferenceImages": [("images", "IMAGE"), ("count", "INT")],
    "StyleRefLogin": [("status", "STRING")],
    "CheckpointLoaderSimple": [("MODEL", "MODEL"), ("CLIP", "CLIP"), ("VAE", "VAE")],
    "FluxGuidance": [("CONDITIONING", "CONDITIONING")],
    "UNETLoader": [("MODEL", "MODEL")],
    "CLIPLoader": [("CLIP", "CLIP")],
    "VAELoader": [("VAE", "VAE")],
    "CLIPTextEncode": [("CONDITIONING", "CONDITIONING")],
    "ConditioningZeroOut": [("CONDITIONING", "CONDITIONING")],
    "EmptyLatentImage": [("LATENT", "LATENT")],
    "EmptySD3LatentImage": [("LATENT", "LATENT")],
    "KSampler": [("LATENT", "LATENT")],
    "VAEDecode": [("IMAGE", "IMAGE")],
    "LoadImage": [("IMAGE", "IMAGE"), ("MASK", "MASK")],
    "SaveImage": [],
    "PreviewImage": [],
    "PreviewAny": [],
    "Note": [],
}

INPUTS = {
    "StyleRefApply": [("style", "STYLEREF_STYLE")],
    "StyleRefFacets": [("style", "STYLEREF_STYLE")],
    "StyleRefReferenceImages": [("style", "STYLEREF_STYLE")],
    "CLIPTextEncode": [("clip", "CLIP"), ("text", "STRING")],
    "ConditioningZeroOut": [("conditioning", "CONDITIONING")],
    "FluxGuidance": [("conditioning", "CONDITIONING")],
    "KSampler": [
        ("model", "MODEL"),
        ("positive", "CONDITIONING"),
        ("negative", "CONDITIONING"),
        ("latent_image", "LATENT"),
    ],
    "VAEDecode": [("samples", "LATENT"), ("vae", "VAE")],
    "SaveImage": [("images", "IMAGE")],
    "PreviewImage": [("images", "IMAGE")],
    # PreviewAny accepts anything — ComfyUI's wildcard type.
    "PreviewAny": [("source", "*")],
}

# ── layout constants ─────────────────────────────────────────────────────────

NODE_WIDTH = 320
V_GAP = 56             # vertical gap between stacked nodes
COL_W = 380            # column pitch; > NODE_WIDTH so columns never touch
COLLAPSED_H = 40       # a collapsed node renders as a title bar only

# Minimum on-canvas height per node type, **measured** in the ComfyUI frontend
# (1.45.x) by declaring each node deliberately too small and reading back the
# size the frontend clamped it to.
#
# This table replaced an estimate built from widget counts, which is worth
# knowing about because of how it failed. The frontend clamps a node *up* to its
# own minimum on load, so under-estimating a height does not produce a small
# node — it produces a node that grows on open and swallows whatever the layout
# parked below it. StyleRefLoad (estimated 248, actually 348) is why the loader
# and the Load node overlapped in every template.
#
# Re-measure after adding or removing a widget on a StyleRef node.
MIN_H = {
    "CheckpointLoaderSimple": 128,
    "UNETLoader": 108,
    "CLIPLoader": 140,
    "VAELoader": 80,
    "CLIPTextEncode": 116,
    "ConditioningZeroOut": 48,
    "FluxGuidance": 80,
    "EmptySD3LatentImage": 144,
    "EmptyLatentImage": 144,
    "KSampler": 312,
    "VAEDecode": 72,
    "SaveImage": 84,
    "PreviewImage": 48,
    "PreviewAny": 152,
    "LoadImage": 128,
    "Note": 92,
    "StyleRefLoad": 348,
    "StyleRefApply": 380,
    "StyleRefFacets": 624,
    "StyleRefReferenceImages": 128,
    "StyleRefLogin": 156,
}

# Node types shipped collapsed. Each is plumbing the user never edits: the text
# encoder is driven by a link (its own text widget is not even used), the
# zero-out has no widgets at all, and VAE Decode is a two-wire junction. Left
# expanded they are three tall boxes of nothing between Apply and the sampler.
#
# The layout reserves only COLLAPSED_H for these, so expanding one on the canvas
# can overlap the node beneath it — the right trade for a canvas that reads
# clearly on open, which is the state every user starts in.
COLLAPSED = {"CLIPTextEncode", "ConditioningZeroOut", "VAEDecode"}

# Node titles longer than this truncate in both frontends.
MAX_TITLE = 28

# Hard cap on a Note node's text. Notes had grown into essays — model setup, a
# no-negative explanation, a guidance-tuning tip — which made the brown box the
# largest thing on the canvas and buried the one or two lines a first-time user
# actually needs. A note now says what the template produces and how to get the
# model; everything else lives in the README and docs.styleref.io.
MAX_NOTE = 340

# The gallery style each set loads. Deliberately different per model, because
# "which style demos well" is a property of the pair, not of the style.
#
# Z-Image renders a painterly style faithfully — panel texture, contour-led
# figures, a muted Renaissance palette — so it gets the painterly one.
#
# FLUX does not. It has a strong photoreal house look and reverts to it,
# and it is specifically weak at reproducing painterly technique; the usual
# remedy in the ecosystem is a style LoRA, which is far too much apparatus for
# a first-run demo. Rather than ship a demo that quietly under-delivers, the
# FLUX set uses a style whose identity is *light and colour* rather than
# brushwork — high contrast, saturated theatrical key, documentary framing.
# That is what FLUX is genuinely good at, and it lands hard enough that nobody
# has to wonder whether the node did anything.
#
# Both are subject-agnostic (the consistency grid needs that) and both carry
# inspiration images (the reference-images template needs that).
#
# Two styles that read well on the gallery still demo badly anywhere here: a
# naturalistic photography style lands close to what a base model does
# unprompted, and a typography-led graphic style asks the sampler for headline
# text. Neither of these does either.
FLUX_STYLE_REF = "azch4ptx-2ca7d3a07f37"    # Raw Theatrical Cinematic
ZIMAGE_STYLE_REF = "sbdlwwly-66ae1a89efdc"  # Renaissance Mythic Classicism

# Placeholder ref for the "your own style" templates. Refs are opaque ids now
# (no name lookup), so the templates ship an obvious placeholder the user swaps
# for their own style's id — or fills via the node's "Search styles…" picker.
OWN_STYLE_REF = "<your style id>"

# Subjects for the consistency grid. Deliberately varied — exterior at dusk,
# interior, and an object with no figure at all — because the grid's whole
# claim is that one style holds across unrelated subjects.
#
# One list per model, for the same reason the style refs differ: what demos
# well is a property of the model/style pair, not of the subject.
#
# Two of the Z-Image subjects say "dressed" outright. That is not prudishness:
# the Renaissance style compiles an empty negative (it carries no guardrails),
# and Z-Image has no negative channel to put one in, so a style steeped in
# classical painting drifts to classical nudes unless the subject says
# otherwise. Positive phrasing in the subject is the only lever that model
# leaves us. FLUX needs none of that — Raw Theatrical Cinematic carries its own
# guardrails and FLUX has a channel for them — so its subjects are chosen
# instead for what the style is actually about: hard cast shadows, a saturated
# key, narrative props and "staged chaos" (spilled cherries and juice in the
# fabric, a fridge door throwing the only fill, a sprinkler on a flat lawn).
ZIMAGE_GRID_SUBJECTS = [
    "A fully dressed teenager sitting on a rooftop ledge at dusk, dress in "
    "#8A9FB5 color with ropes, knees up, headphones on, looking out over a "
    "low-rise city as lights begin to switch on; laundry lines and a water "
    "tank frame the foreground.",
    "two friends flipping through crates of vinyl in a record shop, dressed",
    "A Unicorn Toy",
]

FLUX_GRID_SUBJECTS = [
    "A picnic blanket spread out across grass, covered with a half-eaten "
    "watermelon slice, a tipped bowl of cherries spilling across the stripes, "
    "a transistor radio, sunglasses, and a paperback splayed face-down. Seeds "
    "and rind scattered between them, juice soaked dark into the fabric.",
    "A woman in a robe and hair curlers sitting on a kitchen floor with her "
    "back against the cabinets, eating spaghetti straight from the pot. A "
    "wooden spoon on the tile beside her, an open fridge door casting into "
    "frame.",
    "A woman in a yellow raincoat and boots holding an open umbrella, standing "
    "on a bright green lawn under a garden sprinkler, cloudless sky above. A "
    "watering can and a pair of sandals on the grass beside her.",
]

# The single-subject templates use their own set's first grid subject, so a
# user who opens two templates for the same model sees the same prompt render
# consistently.
ZIMAGE_DEMO_SUBJECT = ZIMAGE_GRID_SUBJECTS[0]
FLUX_DEMO_SUBJECT = FLUX_GRID_SUBJECTS[0]

# Frontend model-download metadata: a missing checkpoint becomes a guided
# download instead of a red error.
#
# The Comfy-Org all-in-one fp8 repack — UNET + CLIP/T5 + VAE in one file, so
# CheckpointLoaderSimple can load it. BFL's own flux1-dev.safetensors is
# UNET-only and cannot.
#
# Why FLUX rather than SDXL, which more people have installed: StyleRef
# compiles a long style spec (~480 tokens for a full style), and SDXL's CLIP
# attends to ~77 tokens per chunk — the style is mostly truncated before it
# reaches the sampler, whatever the checkpoint. FLUX reads the prompt with T5,
# which takes the whole thing. This is the first model where the compiled
# style and the text encoder actually fit each other.
#
# `flux1-schnell-fp8.safetensors` (Comfy-Org/flux1-schnell) is a drop-in swap:
# Apache-2.0 instead of dev's non-commercial licence, and ~4 steps instead of
# 20, at some cost in fidelity.
MODEL_FLUX = {
    "name": "flux1-dev-fp8.safetensors",
    "url": "https://huggingface.co/Comfy-Org/flux1-dev/resolve/main/flux1-dev-fp8.safetensors",
    "directory": "checkpoints",
}

# The sections the Facets templates surface on canvas. Facets has 24 outputs;
# previewing all of them would be a wall of nodes, so the templates show a
# representative six and the note points at the rest. Names are looked up in
# the node's own RETURN_NAMES, so a schema reorder can never mis-slot them.
FACET_SHOWCASE = [
    "colors",
    "light_shadow",
    "mood_personality",
    "surface_material",
    "shape_language",
    "guardrails",
]

# Z-Image Turbo (Tongyi-MAI) — an S3-DiT model with its own loader stack (a
# UNET, a Qwen-3-4B text encoder, and a Flux-style 16-channel VAE), not a
# single checkpoint. URLs are the Comfy-Org repacks. It compiles well from the
# `diffusion` target: qwen handles the tag-weighted output that SDXL/SD1.5 use.
MODEL_ZIMAGE_UNET = {
    "name": "z_image_turbo_bf16.safetensors",
    "url": "https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/diffusion_models/z_image_turbo_bf16.safetensors",
    "directory": "diffusion_models",
}
MODEL_ZIMAGE_CLIP = {
    "name": "qwen_3_4b.safetensors",
    "url": "https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/text_encoders/qwen_3_4b.safetensors",
    "directory": "text_encoders",
}
MODEL_ZIMAGE_VAE = {
    "name": "ae.safetensors",
    "url": "https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/vae/ae.safetensors",
    "directory": "vae",
}
MODELS_ZIMAGE = [MODEL_ZIMAGE_UNET, MODEL_ZIMAGE_CLIP, MODEL_ZIMAGE_VAE]


def node_height(node_type: str) -> int:
    """
    The height a node actually occupies: its measured minimum. Anything smaller
    is overwritten by the frontend on load; anything larger is empty space.
    """
    if node_type not in MIN_H:
        raise KeyError(f"no measured height for {node_type} — add it to MIN_H")
    return MIN_H[node_type]


def effective_size(node: dict[str, Any]) -> tuple[int, int]:
    """
    The box a node occupies on canvas as opened, which is what "do these two
    nodes overlap" has to be asked about. A collapsed node draws as a title bar
    regardless of the `size` it carries for when it is expanded.

    Shared with the layout test so both answer the question the same way.
    """
    if node.get("flags", {}).get("collapsed"):
        return NODE_WIDTH, COLLAPSED_H
    width, height = node["size"]
    return width, height


class Graph:
    """
    Workflow builder with column layout.

    `place(node_type, col, …)` stacks nodes top-to-bottom within a column,
    spacing them by each node's estimated height plus V_GAP — overlap-free by
    construction. Columns are indexed (0, 1, 2, …) and spaced COL_W apart.
    """

    def __init__(self) -> None:
        self.nodes: list[dict[str, Any]] = []
        self.links: list[list[Any]] = []
        self._node_id = 0
        self._link_id = 0
        self._cursor: dict[int, int] = {}  # column index → next free y

    def place(
        self,
        node_type: str,
        col: int,
        widgets: list[Any] | None = None,
        title: str | None = None,
        height: int | None = None,
        models: list[dict[str, str]] | None = None,
    ) -> int:
        if title is not None and len(title) > MAX_TITLE:
            raise ValueError(f"title longer than {MAX_TITLE} chars truncates: {title!r}")

        widgets = widgets or []
        h = height if height is not None else node_height(node_type)
        collapsed = node_type in COLLAPSED
        x = 40 + col * COL_W
        y = self._cursor.get(col, 60)
        # Collapsed nodes only ever draw a title bar, so that is all the column
        # spends on them.
        self._cursor[col] = y + (COLLAPSED_H if collapsed else h) + V_GAP

        self._node_id += 1
        node: dict[str, Any] = {
            "id": self._node_id,
            "type": node_type,
            "pos": [x, y],
            "size": [NODE_WIDTH, h],
            "flags": {"collapsed": True} if collapsed else {},
            "order": len(self.nodes),
            "mode": 0,
            "inputs": [
                {"name": name, "type": kind, "link": None}
                for name, kind in INPUTS.get(node_type, [])
                # `text` on CLIPTextEncode is a widget until something connects
                # to it; declaring it as an input is what lets Apply drive it.
            ],
            "outputs": [
                {"name": name, "type": kind, "links": [], "slot_index": index}
                for index, (name, kind) in enumerate(OUTPUTS.get(node_type, []))
            ],
            "properties": {"Node name for S&R": node_type},
            "widgets_values": widgets,
        }
        if models:
            # ComfyUI reads download metadata from the *node's* properties —
            # this is what turns a missing file into a download offer instead
            # of a silently-substituted combo value. The workflow-level
            # "models" key alone does not drive that prompt.
            node["properties"]["models"] = models
        if title:
            node["title"] = title
        self.nodes.append(node)
        return self._node_id

    # A note wraps at roughly this many characters at NODE_WIDTH.
    NOTE_WRAP = 46

    def note(self, col: int, text: str, title: str = "Read me") -> int:
        """
        A visible core Note node — the instruction channel, and deliberately a
        small one. `MAX_NOTE` caps the text; the height is sized to the wrapped
        line count so the note is exactly as tall as it needs to be.
        """
        if len(text) > MAX_NOTE:
            raise ValueError(
                f"note is {len(text)} chars, over the {MAX_NOTE}-char cap — say what "
                f"this template does and how to get a model, and put the rest in "
                f"the README:\n{text}"
            )
        lines = sum(max(1, -(-len(line) // self.NOTE_WRAP)) for line in text.split("\n"))
        h = max(MIN_H["Note"], 46 + 20 * lines)
        return self.place("Note", col, [text], title=title, height=h)

    def link(self, src: int, src_slot: int, dst: int, dst_slot: int) -> None:
        src_node = self._node(src)
        dst_node = self._node(dst)
        kind = src_node["outputs"][src_slot]["type"]

        self._link_id += 1
        self.links.append([self._link_id, src, src_slot, dst, dst_slot, kind])
        src_node["outputs"][src_slot]["links"].append(self._link_id)
        dst_node["inputs"][dst_slot]["link"] = self._link_id

    def _node(self, node_id: int) -> dict[str, Any]:
        return next(n for n in self.nodes if n["id"] == node_id)

    def to_json(self, models: list[dict[str, str]] | None = None) -> dict[str, Any]:
        payload: dict[str, Any] = {
            "last_node_id": self._node_id,
            "last_link_id": self._link_id,
            "nodes": self.nodes,
            "links": self.links,
            "groups": [],
            "config": {},
            "extra": {"styleref": {"notes": PROVENANCE}},
            "version": 0.4,
        }
        if models:
            payload["models"] = models
        return payload


# Shared widget rows (single source for the values the nodes expect, in the
# INPUT_TYPES order of each node).
def load_widgets(style_ref: str) -> list[Any]:
    # style_ref, search, category, use_cache, save_to_library
    return [style_ref, "", "any", True, False]


def apply_widgets(subject: str, target: str) -> list[Any]:
    # subject, target, extra_negative, compact, sections, subject_first
    # Targets mirror the web app's copy list: diffusion drives SDXL/SD1.5
    # checkpoints, flux drives FLUX ones.
    return [subject, target, "", False, "", True]


def _flux_checkpoint(graph: Graph) -> int:
    """
    Place the FLUX checkpoint loader with its download metadata attached.

    A helper rather than a bare `place()` call for the same reason
    `_zimage_loaders` is one: the metadata has to travel with the node. When
    declaring the download was a separate step, every template placed the
    loader correctly and none of them offered the download.
    """
    return graph.place(
        "CheckpointLoaderSimple", 0, [MODEL_FLUX["name"]], models=[MODEL_FLUX]
    )


def flux_sampler_widgets() -> list[Any]:
    """
    FLUX dev defaults: 20 steps, CFG 1.0, euler/simple.

    CFG stays at 1.0 for the same reason it does on Z-Image — FLUX is guidance
    distilled, and prompt strength is set with FluxGuidance instead. Raising CFG
    here does not sharpen the style, it just doubles the work and burns the
    image.
    """
    # Seeds randomize: a fixed seed makes the second queue a silent cache
    # no-op — deadly in a demo.
    return [0, "randomize", 20, 1.0, "euler", "simple", 1.0]


def _flux_tail(graph: Graph, apply_node: int, checkpoint: int, col: int) -> None:
    """
    Apply → CLIP encoder → FluxGuidance → KSampler → VAEDecode → SaveImage.

    **No negative encoder**, the same story as `_zimage_tail`: at CFG 1.0 the
    negative branch is multiplied out entirely, so a second encoder would be
    dead weight. The sampler's negative input takes a zeroed conditioning.
    Apply's `negative` output stays wired to its Preview Any so the compiled
    exclusions remain readable; fold anything you need from them into `subject`
    as a positive phrase.

    The latent is `EmptySD3LatentImage`, not `EmptyLatentImage`: FLUX works in a
    16-channel latent space and the plain node emits 4 channels. This graph
    previously shipped with the 4-channel node — it is the kind of mistake that
    surfaces as a shape error or noise at sample time, never at build time.
    """
    positive = graph.place("CLIPTextEncode", col, [""], title="Positive (from StyleRef)")
    # FLUX's own default. It briefly shipped at 5.0 here, to force a painterly
    # style that FLUX was never going to reproduce; with a style the model is
    # actually good at, that compensation stops helping and starts costing the
    # subject. Guidance is a blunt "obey harder" dial, and this style already
    # asks for a saturated key and hard shadows — over-driven, the colour goes
    # poster-flat and the staged clutter collapses into contrast. Style
    # adherence is the pairing's job, not this number's.
    guidance = graph.place("FluxGuidance", col, [3.5], title="Flux guidance")
    zero_out = graph.place("ConditioningZeroOut", col, [], title="FLUX has no negative")
    latent = graph.place("EmptySD3LatentImage", col, [1024, 1024, 1])
    sampler = graph.place("KSampler", col + 1, flux_sampler_widgets())
    decode = graph.place("VAEDecode", col + 2)
    save = graph.place("SaveImage", col + 2, ["StyleRef"])

    graph.link(apply_node, 0, positive, 1)
    graph.link(checkpoint, 1, positive, 0)
    graph.link(positive, 0, guidance, 0)
    graph.link(positive, 0, zero_out, 0)
    graph.link(checkpoint, 0, sampler, 0)
    graph.link(guidance, 0, sampler, 1)
    graph.link(zero_out, 0, sampler, 2)
    graph.link(latent, 0, sampler, 3)
    graph.link(sampler, 0, decode, 0)
    graph.link(checkpoint, 2, decode, 1)
    graph.link(decode, 0, save, 0)


def _facet_board(graph: Graph, load: int, col: int) -> int:
    """
    Load → Facets, with a Preview Any on each showcased section.

    Facets outputs plain STRINGs, so the only thing a template can *show*
    without depending on third-party packs is the text itself — which is the
    point: seeing `colors` as readable lines is what tells a user they can drag
    that output into any text input downstream. Returns the Facets node id so
    callers can wire more sections if they want.
    """
    facets = graph.place("StyleRefFacets", col, [], title="Facets — every section")
    graph.link(load, 0, facets, 0)

    for section in FACET_SHOWCASE:
        slot = StyleRefFacets.RETURN_NAMES.index(section)
        preview = graph.place(
            "PreviewAny", col + 1, [], title=f"Facet — {section.replace('_', ' ')}"
        )
        graph.link(facets, slot, preview, 0)

    return facets


# No template attaches Preview Any nodes to Apply's compiled prompts. They were
# there to prove StyleRef produced something, but Apply prints the composed
# prompt and its token estimate in its own node body (`ui.text`) after a queue —
# so the previews duplicated that, and in the consistency grid they duplicated it
# six times, two tall boxes per subject. Preview Any survives only in the Facets
# templates, where the strings have nowhere else to be read.


def zimage_sampler_widgets() -> list[Any]:
    """
    Z-Image Turbo defaults: ~8 steps, CFG 1.0, euler/simple.

    CFG stays at 1.0 because Z-Image has no negative channel to activate (see
    `_zimage_tail`) — raising it only oversaturates and slows the sample down.
    """
    # Seeds randomize: a fixed seed makes the second queue a silent cache
    # no-op — deadly in a demo.
    return [0, "randomize", 8, 1.0, "euler", "simple", 1.0]


def _zimage_loaders(graph: Graph) -> tuple[int, int, int]:
    """Place the Z-Image UNET + CLIP + VAE loaders. Returns their node ids."""
    unet = graph.place(
        "UNETLoader",
        0,
        [MODEL_ZIMAGE_UNET["name"], "default"],
        title="Z-Image UNET",
        models=[MODEL_ZIMAGE_UNET],
    )
    clip = graph.place(
        "CLIPLoader",
        0,
        [MODEL_ZIMAGE_CLIP["name"], "lumina2", "default"],
        title="Qwen text encoder",
        models=[MODEL_ZIMAGE_CLIP],
    )
    vae = graph.place(
        "VAELoader",
        0,
        [MODEL_ZIMAGE_VAE["name"]],
        title="Z-Image VAE",
        models=[MODEL_ZIMAGE_VAE],
    )
    return unet, clip, vae


def _zimage_tail(
    graph: Graph, apply_node: int, unet: int, clip: int, vae: int, col: int
) -> None:
    """
    Apply → one CLIP encoder → KSampler → VAEDecode → SaveImage, on the Z-Image
    loader stack.

    **No negative encoder.** Z-Image's official pipeline runs without
    classifier-free guidance — the authors state the model does not use negative
    prompts at all — so a second encoder here would be dead weight wired to a
    sampler input that cannot read it. The sampler's negative gets a zeroed
    conditioning instead, the same shape a FLUX graph uses. Apply's `negative`
    output stays connected to its Preview Any so the compiled exclusions are
    still readable and copyable; anything you need from them belongs in
    `subject` on Apply.
    """
    positive = graph.place("CLIPTextEncode", col, [""], title="Positive (from StyleRef)")
    zero_out = graph.place("ConditioningZeroOut", col, [], title="Z-Image has no negative")
    latent = graph.place("EmptySD3LatentImage", col, [1024, 1024, 1])
    sampler = graph.place("KSampler", col + 1, zimage_sampler_widgets())
    decode = graph.place("VAEDecode", col + 2)
    save = graph.place("SaveImage", col + 2, ["StyleRef"])

    graph.link(apply_node, 0, positive, 1)
    graph.link(clip, 0, positive, 0)
    graph.link(unet, 0, sampler, 0)
    graph.link(positive, 0, sampler, 1)
    graph.link(positive, 0, zero_out, 0)
    graph.link(zero_out, 0, sampler, 2)
    graph.link(latent, 0, sampler, 3)
    graph.link(sampler, 0, decode, 0)
    graph.link(vae, 0, decode, 1)
    graph.link(decode, 0, save, 0)


# One line each on where the model comes from — the only setup fact a note has
# to carry, because it is the only one that stops a first queue from working.
# Sampler defaults, the no-negative story and the style-strength dials are in the
# README (github.com/StyleRef/styleref-comfyui) and docs.styleref.io.
_ZIMAGE_SETUP_NOTE = (
    "First queue downloads 3 files: UNET, Qwen text encoder, VAE."
)

_FLUX_SETUP_NOTE = (
    "First queue downloads FLUX.1 dev fp8 (~16 GB) to checkpoints/."
)

_MORE = "Sampler tips + stronger styles: see the README."


def workflow_zimage_01_quickstart() -> dict[str, Any]:
    """Z-Image quickstart: Load a gallery style → Apply → generate."""
    g = Graph()
    g.note(
        0,
        "One gallery style, one image, on Z-Image Turbo. No account needed.\n"
        "Queue it. Then edit `subject` on Apply — or swap `style_ref` on Load "
        "for any gallery slug — and queue again.\n" + _ZIMAGE_SETUP_NOTE,
        title="Z-Image quickstart",
    )
    load = g.place(
        "StyleRefLoad", 0, load_widgets(ZIMAGE_STYLE_REF), title="Load — swap this slug"
    )
    unet, clip, vae = _zimage_loaders(g)
    apply_node = g.place(
        "StyleRefApply", 1, apply_widgets(ZIMAGE_DEMO_SUBJECT, "diffusion")
    )
    g.link(load, 0, apply_node, 0)
    _zimage_tail(g, apply_node, unet, clip, vae, col=2)

    return g.to_json(models=MODELS_ZIMAGE)


def workflow_zimage_02_consistency() -> dict[str, Any]:
    """One style across three subjects on Z-Image."""
    g = Graph()
    g.note(
        0,
        "One style, three subjects, three Z-Image samplers. No account needed.\n"
        "Queue once and compare the set. Edit any subject, or swap the style on "
        "Load, and queue again.\n" + _ZIMAGE_SETUP_NOTE,
        title="Z-Image consistency",
    )
    load = g.place("StyleRefLoad", 0, load_widgets(ZIMAGE_STYLE_REF))
    unet, clip, vae = _zimage_loaders(g)

    subjects = ZIMAGE_GRID_SUBJECTS
    for index, subject in enumerate(subjects):
        apply_node = g.place(
            "StyleRefApply", 1, apply_widgets(subject, "diffusion"), title=f"Subject {index + 1}"
        )
        g.link(load, 0, apply_node, 0)
        _zimage_tail(g, apply_node, unet, clip, vae, col=2)

    return g.to_json(models=MODELS_ZIMAGE)


def workflow_zimage_03_your_own_style() -> dict[str, Any]:
    """Your own style, made on the web, rendered on Z-Image."""
    g = Graph()
    g.note(
        0,
        "Your own style on Z-Image. Extract one from any image at styleref.io.\n"
        "1. Sign in with the Login node's button.\n"
        "2. Put your style's id in `style_ref` on Load, or pick it with "
        "Search styles….\n"
        "3. Queue.\n" + _ZIMAGE_SETUP_NOTE,
        title="Z-Image — your style",
    )
    login = g.place("StyleRefLogin", 0, [], title="Login — check status")
    assert login
    load = g.place(
        "StyleRefLoad", 0, load_widgets(OWN_STYLE_REF), title="Load — your own style"
    )
    unet, clip, vae = _zimage_loaders(g)
    apply_node = g.place(
        "StyleRefApply", 1, apply_widgets(ZIMAGE_DEMO_SUBJECT, "diffusion")
    )
    g.link(load, 0, apply_node, 0)
    _zimage_tail(g, apply_node, unet, clip, vae, col=2)

    return g.to_json(models=MODELS_ZIMAGE)


def workflow_zimage_04_reference_images() -> dict[str, Any]:
    """Z-Image render + the style's inspiration images side by side."""
    g = Graph()
    g.note(
        0,
        "The style's prompt and its own inspiration images, together.\n"
        "Apply drives the sampler; Reference Images pulls the images out as a "
        "batch — preview them, or wire them into IPAdapter or ControlNet.\n"
        "Needs a gallery style that has inspiration images.\n" + _ZIMAGE_SETUP_NOTE,
        title="Z-Image + references",
    )
    load = g.place(
        "StyleRefLoad", 0, load_widgets(ZIMAGE_STYLE_REF), title="Load — a gallery slug"
    )
    unet, clip, vae = _zimage_loaders(g)

    refs = g.place("StyleRefReferenceImages", 1, [12], title="Reference images")
    ref_preview = g.place("PreviewImage", 1, [], title="Inspiration images")
    g.link(load, 0, refs, 0)
    g.link(refs, 0, ref_preview, 0)

    apply_node = g.place(
        "StyleRefApply", 1, apply_widgets(ZIMAGE_DEMO_SUBJECT, "diffusion")
    )
    g.link(load, 0, apply_node, 0)
    _zimage_tail(g, apply_node, unet, clip, vae, col=2)

    return g.to_json(models=MODELS_ZIMAGE)


def workflow_01_quickstart() -> dict[str, Any]:
    """Load a gallery style → Apply → generate. The 60-second demo."""
    g = Graph()
    g.note(
        0,
        "One gallery style, one image, on FLUX.1 dev. No account needed.\n"
        "Queue it. Then edit `subject` on Apply — or swap `style_ref` on Load "
        "for any gallery slug — and queue again.\n"
        + _FLUX_SETUP_NOTE
        + "\n"
        + _MORE,
        title="Quickstart",
    )
    load = g.place(
        "StyleRefLoad",
        0,
        load_widgets(FLUX_STYLE_REF),
        title="Load — swap this slug",
    )
    checkpoint = _flux_checkpoint(g)
    apply_node = g.place(
        "StyleRefApply", 1, apply_widgets(FLUX_DEMO_SUBJECT, "flux")
    )
    g.link(load, 0, apply_node, 0)
    _flux_tail(g, apply_node, checkpoint, col=2)

    return g.to_json(models=[MODEL_FLUX])


def workflow_02_consistency() -> dict[str, Any]:
    """
    One style across three subjects, side by side.

    Consistency is the claim this demonstrates, and it is scoped precisely: the
    same style applied within this one model produces a coherent set. It is not
    a before/after, and not a claim that other tools match this output.
    """
    g = Graph()
    g.note(
        0,
        "One style, three subjects, three FLUX samplers. No account needed.\n"
        "Queue once and compare the set. Edit any subject, or swap the style on "
        "Load, and queue again.\n" + _FLUX_SETUP_NOTE,
        title="Consistency grid",
    )
    load = g.place("StyleRefLoad", 0, load_widgets(FLUX_STYLE_REF))
    checkpoint = _flux_checkpoint(g)

    subjects = FLUX_GRID_SUBJECTS
    for index, subject in enumerate(subjects):
        apply_node = g.place(
            "StyleRefApply", 1, apply_widgets(subject, "flux"), title=f"Subject {index + 1}"
        )
        g.link(load, 0, apply_node, 0)
        _flux_tail(g, apply_node, checkpoint, col=2)

    return g.to_json(models=[MODEL_FLUX])


def workflow_03_your_own_style() -> dict[str, Any]:
    """Your own style, made on the web, loaded here (template 03)."""
    g = Graph()
    g.note(
        0,
        "Your own style on FLUX. Extract one from any image at styleref.io.\n"
        "1. Sign in with the Login node's button.\n"
        "2. Put your style's id in `style_ref` on Load, or pick it with "
        "Search styles….\n"
        "3. Queue.\n" + _FLUX_SETUP_NOTE,
        title="Your own style",
    )
    login = g.place("StyleRefLogin", 0, [], title="Login — check status")
    assert login  # not linked: it is a status helper, not a data node
    load = g.place(
        "StyleRefLoad",
        0,
        load_widgets(OWN_STYLE_REF),
        title="Load — your own style",
    )
    checkpoint = _flux_checkpoint(g)
    apply_node = g.place(
        "StyleRefApply", 1, apply_widgets(FLUX_DEMO_SUBJECT, "flux")
    )
    g.link(load, 0, apply_node, 0)
    _flux_tail(g, apply_node, checkpoint, col=2)

    return g.to_json(models=[MODEL_FLUX])


_FACETS_NOTE = (
    "Inside a style: Facets emits each of its 24 sections as its own STRING.\n"
    "Six are previewed here; drag any output into any text input — a second "
    "encoder, an IPAdapter prompt, a LoRA trigger.\n"
    "Apply and the sampler alongside render the whole assembled prompt."
)


def workflow_04_reference_images() -> dict[str, Any]:
    """FLUX render + the style's inspiration images side by side."""
    g = Graph()
    g.note(
        0,
        "The style's prompt and its own inspiration images, together.\n"
        "Apply drives the sampler; Reference Images pulls the images out as a "
        "batch — preview them, or wire them into IPAdapter or ControlNet.\n"
        "Needs a gallery style that has inspiration images.\n" + _FLUX_SETUP_NOTE,
        title="FLUX + references",
    )
    load = g.place(
        "StyleRefLoad", 0, load_widgets(FLUX_STYLE_REF), title="Load — a gallery slug"
    )
    checkpoint = _flux_checkpoint(g)

    refs = g.place("StyleRefReferenceImages", 1, [12], title="Reference images")
    ref_preview = g.place("PreviewImage", 1, [], title="Inspiration images")
    g.link(load, 0, refs, 0)
    g.link(refs, 0, ref_preview, 0)

    apply_node = g.place(
        "StyleRefApply", 1, apply_widgets(FLUX_DEMO_SUBJECT, "flux")
    )
    g.link(load, 0, apply_node, 0)
    _flux_tail(g, apply_node, checkpoint, col=2)

    return g.to_json(models=[MODEL_FLUX])


def workflow_05_facets() -> dict[str, Any]:
    """Every section of a style, exposed and readable, next to a normal render."""
    g = Graph()
    g.note(0, _FACETS_NOTE, title="Facets")
    load = g.place(
        "StyleRefLoad",
        0,
        load_widgets(FLUX_STYLE_REF),
        title="Load — swap this slug",
    )
    checkpoint = _flux_checkpoint(g)

    _facet_board(g, load, col=1)

    apply_node = g.place(
        "StyleRefApply", 3, apply_widgets(FLUX_DEMO_SUBJECT, "flux")
    )
    g.link(load, 0, apply_node, 0)
    _flux_tail(g, apply_node, checkpoint, col=4)

    return g.to_json(models=[MODEL_FLUX])


def workflow_zimage_05_facets() -> dict[str, Any]:
    """The Facets story on the Z-Image loader stack."""
    g = Graph()
    g.note(0, _FACETS_NOTE + "\n" + _ZIMAGE_SETUP_NOTE, title="Facets on Z-Image")
    load = g.place(
        "StyleRefLoad",
        0,
        load_widgets(ZIMAGE_STYLE_REF),
        title="Load — swap this slug",
    )
    unet, clip, vae = _zimage_loaders(g)

    _facet_board(g, load, col=1)

    apply_node = g.place(
        "StyleRefApply", 3, apply_widgets(ZIMAGE_DEMO_SUBJECT, "diffusion")
    )
    g.link(load, 0, apply_node, 0)
    _zimage_tail(g, apply_node, unet, clip, vae, col=4)

    return g.to_json(models=MODELS_ZIMAGE)


# Keyed by path relative to workflows/. The FLUX set lives at the root; the
# Z-Image set lives in workflows/z-image/.
WORKFLOWS = {
    "01-quickstart.json": workflow_01_quickstart,
    "02-consistency-grid.json": workflow_02_consistency,
    "03-your-own-style.json": workflow_03_your_own_style,
    "04-reference-images.json": workflow_04_reference_images,
    "05-facets.json": workflow_05_facets,
    "z-image/01-quickstart.json": workflow_zimage_01_quickstart,
    "z-image/02-consistency-grid.json": workflow_zimage_02_consistency,
    "z-image/03-your-own-style.json": workflow_zimage_03_your_own_style,
    "z-image/04-reference-images.json": workflow_zimage_04_reference_images,
    "z-image/05-facets.json": workflow_zimage_05_facets,
}


def main() -> None:
    for relpath, builder in WORKFLOWS.items():
        path = os.path.join(WORKFLOW_DIR, relpath)
        os.makedirs(os.path.dirname(path), exist_ok=True)
        with open(path, "w", encoding="utf-8") as fh:
            json.dump(builder(), fh, indent=2)
            fh.write("\n")
        print(f"wrote {path}")


if __name__ == "__main__":
    main()