Neons Style Explorer

Every checkpoint has a style vocabulary and almost none of it is documented. The anime bases solved this with artist-tag explorers, and the advice there is consistent: treat style discovery as setup work, find a dozen looks you like, prompt from those. Neons Style Explorer is that idea as a ComfyUI node - 3,015 style entries, a dice to roll them, and a batch mode that generates previews of the lot.

What it actually is

Text in, text out. It renders nothing, touches no sampler, and sits in front of your CLIP Text Encode handing over a prompt you didn't write. Each entry is a style prefix and nothing else - how the picture is rendered, closing with its medium: ... anime style image, ... photograph style image. No subject, no instructions, no fighting your text. The pack lints for that rule.

The catalog is 1,419 hand-written clauses plus 1,596 imported from ThetaCursed's Krea 2 style collection (MIT, credited in THIRD-PARTY-NOTICES.md), across nine families: anime & manga, photography & film, traditional painting, 3D & games, illustration, comics & print, design & aesthetics, experimental & material, western animation. Then 97 formats and 51 finishes on separate axes.

How the composing works

One composer does the work, and the node's own panel calls it over HTTP - so what you see previewed is exactly the string you'll get. With style_position on start:

masterpiece, best quality, Anime broadcast still: flat cel colouring in hard shadow bands, simplified on-model faces, painted matte background, mild compression softness and grain, anime style image. a woman on a fire escape at night

quality always sits at the very front and your prompt follows the clause. Only the leading style contributes a closing medium - extras get theirs stripped, so you never get two mediums in one prompt. Stack three via style / style_2 / style_3, joined by style_mix: blended with, mixed with, layered over, then also. custom_style takes your own entries and stays empty until you write one.

style_weight surprises people. 1.0 is neutral; above it, natural mode wraps the clause as (clause:1.5) and the booru modes wrap every tag as (tag:1.5). Colons inside the clause are rewritten to hyphens first so CLIP doesn't read them as nested weights. output_format picks natural, danbooru or natural + danbooru, tag_separator applies in the booru modes only, and include_style_negative (on by default) merges the style's own avoid terms into your negative.

The controls that matter

• prompt, quality, negative - scene, the masterpiece, best quality box, your avoid terms. Put quality tags in quality so they stop drifting.
• style - the dice lives at the top. roll_scope sets the pool (everything, the primary style's family, favourites, recent, or by preview state) and roll_seed seeds it; use the control-after-generate widget, set to randomize. 0 also rolls fresh each run.
• crawl + crawl_missing_only + auto_gallery - the batch machine. Crawl walks the dropdown one entry per queued run, starting where you parked it. crawl_missing_only on, auto_gallery on every, queue a few hundred: the gallery fills itself in order.

Outputs

positive to your positive CLIP Text Encode, negative to the negative one. debug is worth wiring to a Preview Text node while you learn the thing - chosen styles, medium, weight, crawl state, both composed strings.

Install

Manager: search Neons Style Explorer, install, restart ComfyUI. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Neon-Sparks/ComfyUI-NeonsStyleExplorer

Restart the process, not just the browser. The front end talks to the node over HTTP routes registered at startup, so a refresh loads the new interface but not the routes - and half the controls will tell you the server hasn't loaded this version. No extra dependencies: Pillow and numpy, both already in ComfyUI.

Where people get burned

The weighting slider does nothing on LLM-encoded models. Krea 2, Qwen-Image, Anima, the Flux 2 family: attention weights are discarded, so your parentheses and :1.5 arrive as literal punctuation - worse than not weighting. Leave style_weight at 1.0 there and say it in words.

The Extra family reads differently. Those 1,596 entries are flat descriptor lists, not written clauses, and their booru tags are the same descriptors. Filter by family or source when you want only the written catalog.

custom_style stays empty until you create one, and you'll want Refresh Node Definitions after the first.

Three things that look like bugs and aren't. The dropdown moving on its own means crawl is on. Favourites or recents empty makes the dice fall back to the whole catalog rather than doing nothing. A deleted style is hidden, not destroyed - Restore deleted brings it back. And the gallery ships empty; no previews are bundled, that part is on you.