Workflow templates

Five workflows ship with the node pack, in the workflows/ folder of the repository . Drag a .json file onto the ComfyUI canvas to load it.

These run on FLUX.1 dev — the Comfy-Org fp8 all-in-one checkpoint, which downloads on the first queue. FLUX reads the prompt with a T5 encoder, so a full compiled style fits. An SDXL-family CLIP encoder truncates at roughly 77 tokens, which is why a detailed style washes out there whatever checkpoint you load.

They load a theatrical cinematic gallery style — hard cast shadows, a saturated key, staged narrative props. That is a deliberate pairing: FLUX is strongest at light and colour rather than brushwork, and less willing to reproduce painterly technique, where it tends back toward its own look. The Z-Image set loads a painterly Renaissance style instead, for the same reason in reverse. Swap style_ref on Load for any slug from the gallery  — these are starting points, not limits.

Three need no account.

FLUX takes no negative prompt. It is guidance distilled and runs at CFG 1.0, where the negative branch is multiplied out entirely — so these canvases carry no negative encoder, and the sampler’s negative input gets a zeroed conditioning. Prompt strength is the FluxGuidance node (3.5), not CFG. StyleRef still compiles the negative and returns it on Apply’s negative output; put anything you need from it into subject on Apply as a positive phrase instead. The Z-Image set works the same way, for the same reason.

Each template carries a short note on the canvas — what it produces, what to change, and where the model comes from — and declares its model downloads, so a missing file offers a guided download instead of a red error. Seeds are set to randomize, so every queue produces a fresh image.

The canvases are built to be readable the moment they open: the plumbing between Apply and the sampler — the text encoder, the zeroed conditioning, VAE Decode — ships collapsed, because none of it is yours to edit. Expand any node if you want to rewire it. And Apply prints the prompt it composed, plus a token estimate, in its own node body after a queue, so what StyleRef sent the sampler is always readable without extra nodes.

1. Quickstart

File: 01-quickstart.json · No account needed

The minimum useful graph: a gallery style, a subject, a sampler. After the queue, Apply’s node body shows the prompt it composed, so you can read exactly what StyleRef produced.

To run it: queue. The FLUX checkpoint downloads on the first run.

To make it yours: change subject on Apply, and swap style_ref on Load for any slug from the gallery  — or use the Search styles… button.

If you’re new to the nodes, start here — see install & quickstart.

2. Consistency grid

File: 02-consistency-grid.json · No account needed

One Load feeding three Apply nodes with three different subjects, each through its own sampler. Queue once and compare the three results side by side: the style holds steady while the content changes.

To run it: queue. Change any of the three subjects and queue again.

This is the template that shows what a style specification actually buys you: a coherent set from one definition, rather than prompt text you re-tune per image.

What this demonstrates, precisely: consistency within this model, across generations. It is not a before/after comparison, and it does not claim other tools will produce matching images from the same style. Each tool renders its own way — the style is what stays constant in the instructions.

3. Your own style

File: 03-your-own-style.json · Sign-in required

The full loop, with extraction where it belongs — on the web:

1. Extract a style from any image at styleref.io  (upload, preview, refine). It lands in your library.
2. Sign in — the template includes a Login node; queue it alone first if its status says signed out.
3. Put your style’s id in style_ref on Load — or pick it with Search styles…, which fills the field for you. The field ships holding <your style id>, a placeholder: queueing before you replace it just tells you to fill it in.
4. Queue.

Loading and applying your own style in ComfyUI is free — the extraction on the web is the only step that spends credits.

4. Reference images

File: 04-reference-images.json · No account needed

A style carries more than the words it compiles to. Reference Images pulls the style’s inspiration images out as an image batch, previewed next to a normal render of the same style.

The prompt says how the style reads; the images show how it looks. Wire the batch into IPAdapter or a reference-only ControlNet to use both at once.

To run it: queue. Use a gallery style that has inspiration images — not every style does, and the node stops with a message naming the style when it has none.

5. Facets

File: 05-facets.json · No account needed

Apply hands you one prompt string. Facets hands you the style’s sections one output at a time — 24 of them, exactly the sections on the style board at styleref.io, plus the six custom items.

Six sections are wired to Preview nodes here so you can read them on the canvas: colours, light and shadow, mood, surface and material, shape language, and guardrails. Every other output is on the node too — drag any of them wherever you need it.

To run it: queue. Swap style_ref on Load for any gallery slug to inspect a different style.

They’re plain text outputs, 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’s the escape hatch — take the palette without taking the whole prompt.

The template also runs a normal Apply render alongside, so you can compare the assembled prompt against the parts it was built from.

Z-Image templates

A parallel set wired for Z-Image Turbo lives in workflows/z-image/. Z-Image is an S3-DiT model with its own loader stack — a UNET, a Qwen-3-4B text encoder, and a Flux-style VAE — rather than a single checkpoint; the three files download on first run. StyleRef compiles for it with the diffusion target, the same tag-weighted output SDXL and SD1.5 use.

The sampler is set to Z-Image Turbo defaults — 8 steps, CFG 1.0, euler/simple.

Z-Image takes no negative prompt. The model runs without classifier-free guidance, so there is no negative encoder on these canvases — the sampler’s negative input gets a zeroed conditioning, the same shape a FLUX graph uses. StyleRef still compiles the negative and returns it on Apply’s negative output; put anything you need from it into subject on Apply as a positive phrase instead (“fully clothed”, “no text or watermark”). Raising CFG does not bring the negative back — it only oversaturates the image and slows the sample down.

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 own default. 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. sections on Apply — narrow it 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 given style actually carries before you pick sections.

Regenerating the templates

The templates are generated from a script rather than hand-edited, so node and link IDs stay correct:

python scripts/build_workflows.py

Edit scripts/build_workflows.py and re-run it — changes made directly to the JSON are overwritten.