Verbing Styler (Advanced)

VerbingStylerAdvanced injects an action into your prompt - a gerund, an "-ing" verb - from a menu of about 3,500 of them, abandoning, abating, abounding, on and on. The idea is to describe what your subject is doing from a dropdown instead of writing it. It's one of the pack's "grammatical" nodes, and like its adjective and preposition siblings, it's more of a novelty than a workhorse. Useful for exploration and batching odd variations; not something I'd wire into a serious pipeline.

The "Advanced" suffix means it's the SDXL dual-CLIP variant: the verb menu, but with the positive prompt split across SDXL's two encoders and a control for where the negative goes.

How the "Advanced" plumbing works

SDXL uses two text encoders, CLIP-G and CLIP-L. This node feeds them separately - text_positive_g to G, text_positive_l to L - and applies the chosen verb to both, wrapped at the pack's strong default weight (~2.0). MileHighStyler is a fork of the classic SDXL Prompt Styler. negative_prompt_to (Both / G only / L only) routes the entry's built-in negative; Both by default.

The inputs and outputs that matter

• verbing - the ~3,500-entry gerund dropdown; none is off.
• text_positive_g / text_positive_l - your subject per encoder; same text in both if unsure.
• text_negative - your negative, with the style's negative routed by negative_prompt_to.

log_prompt prints the assembled prompt so you can see the verb in context. Six outputs: text_positive_g, text_positive_l, merged text_positive, and text_negative_g / text_negative_l / merged text_negative. Feed _g / _l into a CLIPTextEncodeSDXL, or take the merged text_positive / text_negative into a normal CLIP Text Encode.

Installing it

ComfyUI Manager: search ComfyUI_MileHighStyler, install, restart. Or:

cd ComfyUI/custom_nodes && git clone https://github.com/TripleHeadedMonkey/ComfyUI_MileHighStyler

then restart. No models, no requirements.txt, no dependencies - Python and JSON. Nodes appear under ali1234/stylers.

Where people get burned

• A verb isn't a pose. Injecting running biases the model toward motion; it won't reliably produce a specific, correct action pose. For real pose control you want ControlNet (OpenPose), not a word menu.
• Abstract verbs make noise. Plenty of these 3,500 have no clear visual - the model can't render abrogating. That's the nature of the toy; keep the seed fixed and A/B if you're exploring.
• Advanced when basic would do. The G/L split only pays off with a CLIPTextEncodeSDXL and different per-encoder text. Otherwise use the merged outputs.
• Model era. Dual-CLIP is SDXL-only. Meaningless on the 2026 LLM-encoded models (Flux 2, Z-Image, Qwen) - and on those you'd just write the verb into a sentence anyway, which reads more naturally than a weighted tag.