Neons Style Explorer (Encode)

If you already want the Neons style catalog in your graph, this variant deletes two nodes from it. Same node, same 3,015 entries, same panel - it just does the CLIP encode for you and hands out conditioning instead of text.

What's different from the plain one

It takes one extra required input, clip, and returns positive_conditioning and negative_conditioning alongside the positive, negative and debug strings. The mechanism is what you'd guess: it composes the prompt through the same composer the plain version uses, then encodes both strings with the CLIP object you gave it, so you wire straight into your KSampler and skip the CLIP Text Encode pair entirely.

That's all it is, and that's the argument for it. Four nodes become one, the wiring gets flatter, and the positive and negative string outputs are the exact text that got encoded. When you're generating workflows from a script or hitting the API, that's one node object instead of a subgraph.

The inputs you set

clip is the required one: feed it the same CLIP from your checkpoint loader that would have gone into your text-encode nodes. Wiring a CONDITIONING into it won't work - the socket is CLIP, which is why nothing connects if you're trying to feed an already-encoded prompt in.

Everything else is the standard set. prompt for your subject and scene, quality for the masterpiece, best quality prefix (kept at the very front of the composed text), negative for your avoid terms. style, style_2 and style_3 pick up to three catalog entries, joined in natural language by style_mix; custom_style holds your own written entries and sits empty until you create one. format and finish layer a picture format and a realism qualifier on top. output_format picks natural / danbooru / natural + danbooru, style_position puts the clause before or after your prompt, tag_separator only applies in the booru modes, and include_style_negative merges the style's own avoid terms into negative_conditioning.

style_weight is the emphasis slider - 1.0 is neutral, and above that the clause is wrapped as (clause:1.5) before encoding. Same caveat as the plain node, and it bites harder here because the emphasis is baked into what the encoder sees: on LLM-encoded models like Krea 2, Anima or Qwen-Image, attention weights are discarded, so that syntax arrives as literal punctuation. Keep the slider at 1.0 there and describe the look in words.

The dice and batch controls come along unchanged: style on the 🎲 Random entry, roll_scope for the pool, roll_seed with its control-after-generate widget, and crawl / crawl_missing_only / auto_gallery for the unattended preview run that walks the dropdown one entry per queued prompt.

Outputs

Two conditioning sockets - positive_conditioning to the KSampler's positive, negative_conditioning to the negative - plus three strings. The strings are the reason to keep this node visible rather than collapsing it into a group: positive and negative are the composed prompts, ready for a text display or a Save Text node when you want to log what a run actually sent. debug prints the chosen styles, the medium, the weight, whether crawl was on and both composed strings; wire it to Preview Text for your first afternoon and it'll explain most bad runs on its own.

One thing to know: negative_conditioning is a real encoded negative, not a blank one, because the style's avoid terms are merged in by default. If your workflow wants an empty negative, turn include_style_negative off and clear the negative box.

Install

Same pack, same steps. ComfyUI Manager → search Neons Style Explorer → install → restart ComfyUI. Manually:

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

Both land in the Neons category. Nothing else to install - it runs on the Pillow and numpy ComfyUI already ships, and Python 3.9+ is the only other requirement.

Where people get burned

Restart the server process after installing or updating, not just the browser tab. The node's controls and its catalog routes are registered at startup, so after an update you can watch the new widgets sit there doing nothing until the process comes back. If the panel itself is blank or the layout is scrambled, hard-refresh (Ctrl/Cmd+Shift+R); the front-end files are cached aggressively.

Then one judgement call. Where a model's official workflow routes text through a specialised encoder - the SDXL refiner's conditioning, or an encoder with extra fields like Flux's guidance embedding - the plain Neons Style Explorer plus your own encode node is the safer path, because this one assumes a normal CLIP encode. And don't expect it to speed anything up: it encodes on every run, which is free at this size, but it isn't a cache.