Stylebook Style

"Wet plate collodion" is a fun thing to type and a terrible thing to bet a render on. Stylebook Style stops that bet: 600+ visual styles, each with a rendered preview you can look at before you commit. Open the gallery, pick the look, and the node hands you the prompt text - plus a matching negative - ready to wire into a CLIPTextEncode.

It's the node you start every stylebook workflow from. You type your subject; it adds the style; the two travel down the graph together.

How it works

A Stylebook chain passes state along a style_chain socket, not through the prompt string. This node resolves one style - the medium: photography, painting, anime, print, whatever - and stores it on that chain. The chain is "exclusive" for styles: chain a second Style node and the downstream one wins, with a warning, rather than stacking two mediums nobody meant to stack.

Every style ships four things - a written description, a keyword list, a matching negative prompt, and a preview image - so the node isn't just appending a label, it's choosing how that label gets written.

The two knobs that decide how your model reads it:

• format - prose writes the style as a plain sentence; tags as a comma-separated keyword list. Every style ships both, so switching costs nothing. If your model is LLM-encoded (prompts-as-instructions), prose is your friend; if it's SDXL-lineage tags, tags often lands harder.
• placement - append (the default) puts your subject first and introduces the style after it as a rendering instruction. prepend leads with the style, which is what you want with tags, because a keyword list weights its leading terms most. This ordering is the whole game: lead with a paragraph of film grain and your model draws grain and vaguely gestures at a subject. The node even frames prose styles explicitly ("Rendered as a View-Master stereo slide…"), so the style reads as a process rather than another object in the picture.

You also get strength - subtle keeps only the defining phrase so your subject dominates, normal is the full description, strong repeats the defining term.

The inputs that matter

• mode - defaults to Random, so the node produces a style the moment you hit Run. Pick is for when you use the gallery button and choose one. Cycle steps through the filtered pool by index, which is how you sweep a whole category.
• style - the pick, in Pick mode. Hit Open style gallery to browse with previews instead of guessing.
• category and tag_filter - narrow what Random and Cycle draw from. tag_filter is comma-separated and every term must match (ink, flat finds styles that are both); it searches tags, descriptions, names and aliases, so reverse harem still finds Otome Game.
• seed - for Random mode. Seeds are scored against each candidate rather than indexing a list, so your saved seed keeps producing the same style as the pack grows - roughly one seed in N changes, not all of them.
• user_prompt - your subject, or leave it empty to get style text only. You can connect another pack's text output here (the README wires Identity Forge's subject description in).

The outputs are prompt (full rendered text - wire into a CLIPTextEncode), negative (the style's hand-written negative, for a second encoder), and style_chain (feeds Stylebook Artist and Modifier).

Installing

In ComfyUI Manager, search for Stylebook. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/EnragedAntelope/comfyui-stylebook.git

Restart ComfyUI; the nodes live under conditioning/stylebook. That's the whole install - no third-party dependencies, no model downloads, no API keys, fully offline. One real gotcha: the pack is written against ComfyUI's newer V3 node API, so it needs ComfyUI 0.25.0 or newer (the pyproject.toml says so). If the nodes don't appear, update ComfyUI first - old installs silently won't register a V3 pack.

You can browse every style and its exact prompt in a browser without installing anything at https://enragedantelope.github.io/comfyui-stylebook/docs/gallery/.

Where people get burned

The two classic mistakes are both placement problems. Leading with the style and naming your subject last gets you "the style honoured and the subject ignored" - keep placement on append for prose. And some styles own a subject: Ikebana is an arrangement of flowers, Bonsai is a tree, and they'll assert themselves over your prompt by design. Want the look without the subject? Drop strength to subtle.

If you run at CFG 1 the negative does nothing (by design), so styles defined by what they exclude read softer - most are unaffected. And previews are one model's reading of one style at one seed: a guide, not a guarantee.

Right-click the node for two genuinely useful extras: Copy resolved prompt puts the full untruncated text on your clipboard, and Pin this pick saves whatever Random just landed on into the picker and switches you to Pick mode - the fastest way to keep a style you landed on.