nodes/StyleRef-styleref-comfyui
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
Files
f8e5548b465cb2f6d1ed35b2c2be25d0cd2c3c07
StyleRef-styleref-comfyui/scripts/build_workflows.py
T
BehzadandClaude Sonnet 5 f8e5548b46 Drop internal plan references from code comments 
The pack cited internal planning-doc items (plan P3-7, decision D1, plan
risk 9.2) that aren't public, leaving outside contributors with pointers
they can't follow. Each comment already states its reason; only the
dangling citation is removed.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
2026-07-26 23:18:25 +02:00

730 lines
29 KiB
Python
Raw Blame History

#!/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 estimated from its widget count, and nodes stack per column with a fixed
vertical gap — so no template can ship with overlapping nodes at default sizes.
A test asserts that invariant on the committed files.
Instruction policy: user-facing instructions live in visible core
`Note` nodes on the canvas. `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")],
"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")],
"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
HEADER_H = 60 # title bar
SLOT_H = 24 # one input/output slot row
WIDGET_ROW_H = 28 # one single-line widget
MULTILINE_H = 90 # one multiline text widget
V_GAP = 48 # vertical gap between stacked nodes
COL_W = 380 # column pitch; > NODE_WIDTH so columns never touch
# Widgets that render multiline per node type (by widget index).
MULTILINE_WIDGETS = {
"StyleRefApply": {0, 2}, # subject, extra_negative
"CLIPTextEncode": {0},
"Note": {0},
}
# Node titles longer than this truncate in both frontends.
MAX_TITLE = 28
# Frontend model-download metadata: a missing checkpoint becomes a
# guided download instead of a red error.
# 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>"
MODEL_SDXL = {
"name": "sd_xl_base_1.0.safetensors",
"url": "https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0.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, widgets: list[Any]) -> int:
"""
Estimated on-canvas height: header + one row per input/output
slot + one row per widget. The slot term matters for wide-output nodes —
Facets alone has dozens of outputs.
"""
slots = max(len(INPUTS.get(node_type, [])), len(OUTPUTS.get(node_type, [])))
multiline = MULTILINE_WIDGETS.get(node_type, set())
rows = sum(1 for i in range(len(widgets)) if i not in multiline)
tall = sum(1 for i in range(len(widgets)) if i in multiline)
return HEADER_H + slots * SLOT_H + rows * WIDGET_ROW_H + tall * MULTILINE_H
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,
) -> 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, widgets)
x = 40 + col * COL_W
y = self._cursor.get(col, 60)
self._cursor[col] = y + 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": {},
"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 title:
node["title"] = title
self.nodes.append(node)
return self._node_id
def note(self, col: int, text: str, title: str = "Read me") -> int:
"""A visible core Note node — the instruction channel."""
# Height scales with the text so long notes don't clip at default size.
h = HEADER_H + 30 + 18 * (text.count("\n") + len(text) // 55)
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 sampler_widgets(cfg: float) -> list[Any]:
# Seeds randomize: a fixed seed makes the second queue a silent
# cache no-op — deadly in a demo.
return [0, "randomize", 25, cfg, "euler", "normal", 1.0]
def _sdxl_tail(graph: Graph, apply_node: int, checkpoint: int, col: int) -> None:
"""Apply → two CLIP encoders → KSampler → VAEDecode → SaveImage."""
positive = graph.place("CLIPTextEncode", col, [""], title="Positive (from StyleRef)")
negative = graph.place("CLIPTextEncode", col, [""], title="Negative (from StyleRef)")
latent = graph.place("EmptyLatentImage", col, [1024, 1024, 1])
sampler = graph.place("KSampler", col + 1, sampler_widgets(cfg=7.0))
decode = graph.place("VAEDecode", col + 2)
save = graph.place("SaveImage", col + 2, ["StyleRef"])
graph.link(apply_node, 0, positive, 1)
graph.link(apply_node, 1, negative, 1)
graph.link(checkpoint, 1, positive, 0)
graph.link(checkpoint, 1, negative, 0)
graph.link(checkpoint, 0, sampler, 0)
graph.link(positive, 0, sampler, 1)
graph.link(negative, 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
def _preview(graph: Graph, apply_node: int, col: int, suffix: str = "") -> None:
"""
Compiled prompts made visible, in every template: Preview Any on Apply's
positive and negative kills the "empty text box on the encoder looks
broken" perception and teaches what StyleRef actually produces.
"""
pos = graph.place("PreviewAny", col, [], title=f"Compiled positive{suffix}")
neg = graph.place("PreviewAny", col, [], title=f"Compiled negative{suffix}")
graph.link(apply_node, 0, pos, 0)
graph.link(apply_node, 1, neg, 0)
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.
"""
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"
)
clip = graph.place(
"CLIPLoader", 0, [MODEL_ZIMAGE_CLIP["name"], "lumina2", "default"], title="Qwen text encoder"
)
vae = graph.place("VAELoader", 0, [MODEL_ZIMAGE_VAE["name"]], title="Z-Image 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)
_ZIMAGE_SETUP_NOTE = (
"Z-Image Turbo (S3-DiT). Three model files download on first use: the UNET "
"(diffusion_models/), the Qwen text encoder (text_encoders/), and the VAE "
"(vae/).\n"
"Sampler defaults are turbo: 8 steps, CFG 1.0, euler/simple.\n"
"Z-Image takes no negative prompt. The model runs without classifier-free "
"guidance, so there is no negative encoder on this canvas — the sampler's "
"negative input gets a zeroed conditioning. StyleRef still compiles the "
"negative and previews it; put anything you need from it into `subject` on "
"Apply as a positive phrase (\"fully clothed\", \"no text or watermark\"). "
"Raising CFG does not bring the negative back.\n"
"Style prompts use the `diffusion` target — Z-Image reads the same "
"tag-weighted output SDXL/SD1.5 use."
)
def workflow_zimage_01_quickstart() -> dict[str, Any]:
"""Z-Image quickstart: Load a gallery style → Apply → generate."""
g = Graph()
g.note(
0,
"Z-Image quickstart — no account needed.\n"
"1. Queue — the three Z-Image files download on first run.\n"
"2. Change `subject` on StyleRef Apply and queue again (seeds randomize).\n"
"Swap `style_ref` on StyleRef Load for any gallery slug.\n\n" + _ZIMAGE_SETUP_NOTE,
title="Z-Image quickstart",
)
load = g.place(
"StyleRefLoad", 0, load_widgets("9a2adtz6-cd8d77ee2f51"), title="Load — swap this slug"
)
unet, clip, vae = _zimage_loaders(g)
apply_node = g.place(
"StyleRefApply", 1, apply_widgets("a lighthouse on a rocky shore at dawn", "diffusion")
)
g.link(load, 0, apply_node, 0)
_preview(g, apply_node, 1)
_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,
"Z-Image consistency grid — no account needed.\n"
"One style feeds three Apply nodes with three subjects; each renders "
"through its own Z-Image sampler. Queue once and compare: one style, "
"many subjects, a coherent set within this model.\n\n" + _ZIMAGE_SETUP_NOTE,
title="Z-Image consistency",
)
load = g.place("StyleRefLoad", 0, load_widgets("9a2adtz6-cd8d77ee2f51"))
unet, clip, vae = _zimage_loaders(g)
subjects = [
"a lighthouse on a rocky shore at dawn",
"a quiet street cafe in the rain",
"a hot air balloon over desert dunes",
]
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)
_preview(g, apply_node, 1, suffix=f" {index + 1}")
_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,
"Use your own style on Z-Image.\n"
"Extract a style from any image at https://styleref.io, then load it "
"here by its share slug or id.\n"
"1. Queue the StyleRef Login node alone first (use its Sign in button).\n"
"2. Put your style's id in `style_ref` on StyleRef Load, or pick it\n"
" with the Search styles… button.\n"
"3. Queue.\n\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("a quiet street cafe in the rain", "diffusion")
)
g.link(load, 0, apply_node, 0)
_preview(g, apply_node, 1)
_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,
"Z-Image — style prompt and reference images together.\n"
"StyleRef Apply drives the Z-Image sampler with the style's prompt; "
"StyleRef Reference Images pulls the style's inspiration images out as "
"a batch, previewed alongside. The prompt says how the style reads; the "
"images show how it looks — wire the images into IPAdapter or a "
"reference-only ControlNet to combine both.\n"
"Use a gallery style that has inspiration images.\n\n" + _ZIMAGE_SETUP_NOTE,
title="Z-Image + references",
)
load = g.place(
"StyleRefLoad", 0, load_widgets("9a2adtz6-cd8d77ee2f51"), 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("a lighthouse on a rocky shore at dawn", "diffusion")
)
g.link(load, 0, apply_node, 0)
_preview(g, apply_node, 1)
_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,
"Quickstart — no account needed.\n"
"1. Set your SDXL checkpoint below (a missing model offers a download).\n"
"2. Queue. Seeds randomize, so every queue is a fresh image.\n"
"3. Change `subject` on StyleRef Apply and queue again.\n"
"Swap `style_ref` on StyleRef Load for any slug from "
"https://styleref.io/gallery — or use the Search styles… button.",
title="Quickstart",
)
load = g.place(
"StyleRefLoad",
0,
load_widgets("9a2adtz6-cd8d77ee2f51"),
title="Load — swap this slug",
)
checkpoint = g.place("CheckpointLoaderSimple", 0, [MODEL_SDXL["name"]])
apply_node = g.place(
"StyleRefApply", 1, apply_widgets("a lighthouse on a rocky shore at dawn", "diffusion")
)
_preview(g, apply_node, 1)
g.link(load, 0, apply_node, 0)
_sdxl_tail(g, apply_node, checkpoint, col=2)
return g.to_json(models=[MODEL_SDXL])
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,
"Consistency grid — no account needed.\n"
"One style feeds three Apply nodes with three different subjects; each "
"renders through its own sampler. Queue once and compare: one style, "
"many subjects, a coherent set within this model.\n"
"Change any subject, or swap the style on StyleRef Load, and queue again.",
title="Consistency grid",
)
load = g.place("StyleRefLoad", 0, load_widgets("9a2adtz6-cd8d77ee2f51"))
checkpoint = g.place("CheckpointLoaderSimple", 0, [MODEL_SDXL["name"]])
subjects = [
"a lighthouse on a rocky shore at dawn",
"a quiet street cafe in the rain",
"a hot air balloon over desert dunes",
]
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)
_preview(g, apply_node, 1, suffix=f" {index + 1}")
_sdxl_tail(g, apply_node, checkpoint, col=2)
return g.to_json(models=[MODEL_SDXL])
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,
"Use your own style.\n"
"Extract a style from any image at https://styleref.io — it appears in "
"your library — then load it here by its share slug or id.\n"
"1. Queue the StyleRef Login node alone first (Queue runs every node, "
"so an empty workflow is the clean way). If status says signed out, "
"set action to `sign in` and queue again.\n"
"2. Put your style's id in `style_ref` on StyleRef Load, or pick it\n"
" with the Search styles… button.\n"
"3. Set your SDXL checkpoint and queue.",
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 = g.place("CheckpointLoaderSimple", 0, [MODEL_SDXL["name"]])
apply_node = g.place(
"StyleRefApply", 1, apply_widgets("a quiet street cafe in the rain", "diffusion")
)
g.link(load, 0, apply_node, 0)
_preview(g, apply_node, 1)
_sdxl_tail(g, apply_node, checkpoint, col=2)
return g.to_json(models=[MODEL_SDXL])
_FACETS_NOTE = (
"Inside the style — the Facets node.\n"
"Apply gives you one prompt string. Facets gives you the style's sections "
"one output at a time — 24 of them, exactly the sections you see on the "
"style board at styleref.io, plus the six custom items.\n"
"Six are previewed here so you can read them. Every other output is on the "
"node too; drag any of them into a text input.\n"
"They are plain STRINGs, so they go anywhere text goes: a second CLIP "
"encoder for regional conditioning, an IPAdapter or ControlNet prompt, a "
"LoRA trigger line, or your own nodes. That is the escape hatch — use the "
"palette without taking the whole prompt.\n"
"The render alongside is the paved road (Apply), so you can compare the "
"assembled prompt against the parts it was built from."
)
def workflow_04_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("9a2adtz6-cd8d77ee2f51"),
title="Load — swap this slug",
)
checkpoint = g.place("CheckpointLoaderSimple", 0, [MODEL_SDXL["name"]])
_facet_board(g, load, col=1)
apply_node = g.place(
"StyleRefApply", 3, apply_widgets("a lighthouse on a rocky shore at dawn", "diffusion")
)
g.link(load, 0, apply_node, 0)
_preview(g, apply_node, 3)
_sdxl_tail(g, apply_node, checkpoint, col=4)
return g.to_json(models=[MODEL_SDXL])
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\n" + _ZIMAGE_SETUP_NOTE, title="Facets on Z-Image")
load = g.place(
"StyleRefLoad",
0,
load_widgets("9a2adtz6-cd8d77ee2f51"),
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("a lighthouse on a rocky shore at dawn", "diffusion")
)
g.link(load, 0, apply_node, 0)
_preview(g, apply_node, 3)
_zimage_tail(g, apply_node, unet, clip, vae, col=4)
return g.to_json(models=MODELS_ZIMAGE)
# Keyed by path relative to workflows/. The SDXL 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-facets.json": workflow_04_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()
Reference in New Issue View Git Blame Copy Permalink