StyleRef nodes for ComfyUI

Define a creative style once on StyleRef — colors, lighting, mood, composition, guardrails — then load it as a node and reuse it across your workflows.

One StyleRef Load feeding three Apply nodes with three unrelated subjects — a rooftop at dusk, a record shop, a toy unicorn. The style holds; only the subject changes. That is the whole idea.

The style is compiled server-side for whichever model you point it at, so the same style produces prompt text phrased appropriately for each one.

No account needed to load and apply public gallery styles. An account adds your own private styles. Nothing in this node pack spends credits.

---

Install

ComfyUI Manager (recommended) — search for StyleRef and click Install.

Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/StyleRef/styleref-comfyui

Restart ComfyUI. Nothing to install: the pack declares no Python dependencies and imports on the standard library alone. The Reference Images node reaches for Pillow and numpy when it runs, and every ComfyUI install already ships those. Look for [StyleRef] v1.1.0 — 5 nodes registered in the server console at startup.

It does depend on styleref.io. These nodes are a client, not a local library: every style is fetched from the StyleRef API at https://styleref.io/api/v1, and all prompt compilation happens there. So the pack needs a working internet connection and the service to be up — offline, or during an outage, the nodes fail with an error rather than degrading. Loaded styles are cached for the session, so a graph mid-run keeps working, but the first load of any style is a network call. STYLEREF_API repoints the client if you run your own instance.

Two hosts, then, and no others: the API above, and — only when the Reference Images node runs — the storage host serving that style's inspiration images. Those image requests are deliberately anonymous: your token is never sent to a host outside the API. Nothing here spends credits.

60-second quickstart

1. Add StyleRef Load and paste a slug from the gallery into style_ref (try azch4ptx-2ca7d3a07f37, the Raw Theatrical Cinematic style) — or click Search styles… to pick one.
2. Add StyleRef Apply, connect style, type a subject, pick your target.
3. Wire positive / negative into your CLIP Text Encode nodes.

Or drag workflows/01-quickstart.json onto the canvas.

Nodes

Node	Sign-in	What it does
StyleRef Load	No	Fetch a style by share slug or share URL — and, signed in, your own private styles by id or /styles/{id} URL. Names the loaded style on the node, plus gallery search with a picker dialog and a ↻ Refresh button.
StyleRef Apply	No	Compose a subject with a style into positive / negative prompts for a chosen target, with a prompt preview in the node body.
StyleRef Facets	No	Every schema section as its own output — exactly the style board's section list, plus the six custom style items.
StyleRef Reference Images	No	The style's inspiration images as an IMAGE batch — for IPAdapter and reference conditioning.
StyleRef Login	—	Status line + Sign in / Sign out buttons. Needed only to load your own private styles.

Targets

ai_tools, style_md, flux, midjourney, diffusion, json — exactly the web app's copy-box format list. For the sampler in your graph use flux (FLUX models) or diffusion (SDXL, SD1.5, DALL·E-class); the other four are for copying the output out of ComfyUI, and the node warns if one feeds a CLIP encoder.

Browse the gallery

Every style on styleref.io/gallery loads here by its slug — no account needed.

Making your own styles

Extraction (turning an image into a style) happens on styleref.io — upload an image there, refine the result, and it lands in your library. Load it here by its share slug or id (sign in with the Login node first).

Workflow templates

The templates at the root run on FLUX.1 dev (the Comfy-Org fp8 all-in-one checkpoint, downloaded on first queue). FLUX reads the prompt with T5, so a full compiled style fits — an SDXL CLIP encoder truncates it at ~77 tokens, which is why the style washes out there no matter which checkpoint you use.

File	Sign-in	Story
01-quickstart.json	No	Load → Apply → generate
02-consistency-grid.json	No	One style, three subjects, side by side
03-your-own-style.json	Yes	Extract on the web, load your own style here
04-reference-images.json	No	The style's inspiration images as a batch, next to a render
05-facets.json	No	Every section of the style as its own output, next to a normal render

A parallel set wired for Z-Image Turbo (an S3-DiT model — UNET + Qwen text encoder + VAE, compiled from the diffusion target) lives in workflows/z-image/: the same five stories on that loader stack. The three model files download on first run.

The two sets load different gallery styles on purpose. How well a style demos is a property of the style and the model. FLUX is strongest at light and colour rather than brushwork, so its templates load a theatrical cinematic style — hard cast shadows, a saturated key, staged props. Z-Image reproduces painterly technique faithfully, so its templates load a Renaissance tempera style. Swap either for any slug from the gallery; these are starting points, not limits.

Every template is built to read at a glance on open: the on-canvas notes say what the template produces and where the model comes from, and the plumbing between Apply and the sampler (the text encoder, the zeroed conditioning, VAE Decode) ships collapsed — nothing in there is yours to edit. Expand any of them if you want to rewire.

Neither FLUX nor Z-Image takes a negative prompt. Both are guidance distilled and run at CFG 1.0, where the negative branch is multiplied out entirely — so neither set carries a negative encoder, and the sampler's negative input gets a zeroed conditioning. StyleRef still compiles the negative and returns it on Apply's negative output; fold anything you need from it into subject on Apply as a positive phrase. Raising CFG does not bring it back.

Apply prints the prompt it composed, and its token estimate, in its own node body after a queue — that is where to read what StyleRef actually sent the sampler.

Style not coming through strongly enough?

FLUX has a confident house look and drifts back to it. Two dials, in the order worth trying:

1. Flux guidance (3.5 in the templates, FLUX's default) is the blunt one. 4–5 makes the style bite harder; past that the subject starts losing to it, colour goes poster-flat, and a dark style can go to pure black.
2. Narrow sections on StyleRef Apply to the parts that define this style. For a period or painterly style that is usually artistic_mediums,references,surface_material,colors — the sections naming the medium, the era and the finish. Left empty, every section is sent, and the identity-carrying ones render last, so a long style can read as generic. For a geometric or layout-led style keep shape_language and spatial_hierarchy instead. Queue the Facets template to see what a style actually carries.

Sampler defaults, for reference: FLUX 20 steps, Z-Image Turbo 8, both at CFG 1.0 with euler/simple, seeds randomizing on every queue.

Signing in

Add the StyleRef Login node and click Sign in (browser) — no queueing needed. Your browser opens, you approve, and the node's status flips to signed in. The token is stored at ~/.config/styleref/credentials.json with 0600 permissions. The styleref CLI shares this login.

(No frontend at all? npx styleref login signs in the same credentials file, and queueing the Login node reports the current status.)

Running ComfyUI on a remote or headless machine

Browser sign-in cannot work on a rented GPU box, RunPod, or any headless server: there is no browser to open, and localhost is the wrong machine. Use a token instead:

1. Copy a token from your account page.
2. Where ComfyUI runs, set STYLEREF_TOKEN=<token>
3. Restart ComfyUI.

STYLEREF_TOKEN always takes precedence over stored credentials. The Login node detects the headless case and prints these instructions rather than hanging.

Credits

Nothing in this node pack spends credits. Loading, searching, and applying styles is free and unlimited within the public rate limit. Extraction — the one paid action — happens on styleref.io, where you can preview and refine the result before it costs anything beyond the extraction itself. See plans and limits.

Configuration

Variable	Purpose
STYLEREF_TOKEN	Bearer token; overrides stored credentials. The headless path.
STYLEREF_API	API base URL. Defaults to https://styleref.io/api/v1.
STYLEREF_FORCE_BROWSER_LOGIN	1 forces browser login, 0 forces the token path.

Development

pip install pytest ruff pillow numpy   # pillow/numpy are test-only deps
pytest tests/ -q
ruff check .
python scripts/build_workflows.py   # regenerate workflow templates

The workflow templates are generated — edit scripts/build_workflows.py, not the JSON.

Shipping a change to users — a version bump, a Registry publish, updating the monorepo pointer — is RELEASING.md.

The nodes are a thin client over StyleRef's public REST API v1 (OpenAPI). All prompt compilation happens server-side — this pack never writes style prose itself, so improvements to StyleRef's compiler reach you without an update to this pack.

Docs & support

• ComfyUI connector docs
• Issues

License

MIT — see LICENSE.