🎬 Action

Action is the node that decides what your character is actually doing. The rest of the ComfyUI Scene Composer pack sets the stage - who's in frame, what they're wearing, where they are - and this one picks the pose, the gesture, and, if you flip one toggle, the explicit act. Feed it a seed and a few dropdowns and it hands back a comma-separated string of booru tags to fold into the rest of the scene.

That's the whole pack's trick. Scene Composer, by an author who goes by taches, is a small set of nodes that procedurally build NSFW anime scene prompts so you don't have to hand-type them or babysit a folder of wildcard text files. If you've used dynamic-prompt wildcards, this is the same idea in a node graph instead of a .txt - and Action is the loaded one.

How it works

There's no model and no API here. The node holds curated tag lists internally; your seed plus your dropdown choices deterministically pick from them and format the result. Same seed and same settings give you the same string every time, so you can lock a pose you like and vary everything else, or spin the seed to rifle through options. Set a dropdown to random and it rolls within that category; set it to a specific value and it stays put.

The nsfw boolean is the master switch. With it off (the default), you get a clean, SFW action fragment - a body position and a hand gesture. Flip it on and the explicit dropdowns wake up and get stitched in.

The inputs that matter

• nsfw (boolean) - the gate. Off = SFW pose only; on = the act tags below get added. Set this first, because it decides whether the rest of the node does anything.
• position - the base body pose: random, standing, sitting, lying, walking, running. This is the tag that anchors the figure regardless of the NSFW toggle.
• gesture - what the hands/upper body are doing: random, adjusting hair, waving, hand on hip, crossed arms. Small flavor, but it's the difference between a stiff mannequin and someone caught mid-motion.
• act_type and act - the explicit pair. act_type chooses the category (teasing, preliminaries, intercourses, or random); act is the specific one, a 13-entry list running from mild (flashing, fingering) through to explicit intercourse tags, plus random. Leave act on random and it'll pick within whatever act_type you set.
• cum (boolean) - adds the finishing tag. Off by default.
• seed - reproducibility for every random above.

The single output is a STRING. It's designed to plug into the action input of the Scene node, which merges it with your composition, character, clothes, and environment into one final prompt. You can wire it straight into a CLIP Text Encode, but you'd be throwing away the assembly the pack is built to do.

A thing worth knowing: this is a Pony prompt

Look at the pack's default quality preamble (over on the Composition node) and it's score_9, score_8_up, score_7_up, source_anime - that's the Pony Diffusion V6 XL score-tag string, dead giveaway. This pack speaks booru tags, and booru tags are what Pony and the Illustrious-family SDXL anime models understand. Point the final prompt at a Pony or Illustrious checkpoint and it sings. Point it at Flux, Z-Image, or any 2026 LLM-encoded model and the tags mostly turn to mush - those want plain-language sentences, not a tag pile. Match the node to the right model or nothing else matters.

Installing it

You install the whole pack, not the one node:

• ComfyUI Manager - search ComfyUI Scene Composer, install, restart. Or import the repo's workflow.json and hit Install Missing Nodes.
• Comfy Registry CLI - comfy node registry-install comfyui-scene-composer.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/taches-ai/comfyui-scene-composer, then pip install -r requirements.txt, then restart.

The nodes are pure text generators, so there's no multi-gig model download for the pack itself - but you still need a Pony/Illustrious checkpoint loaded to render anything. One heads-up from the README: v0.3.x was rewritten onto ComfyUI's subgraph system and old workflows break. Start from the repo's own workflow.json rather than something you saved months ago.

Common issues

Getting SFW output when you wanted explicit? Check the nsfw toggle - it's off by default and gates the entire act system. Getting garbage instead of a coherent pose? You're almost certainly on the wrong model family; these tags need an SDXL anime checkpoint. And because this is a niche solo project with basically no community footprint, your best reference isn't a forum thread - it's the author's full NSFW workflow on CivitAI (model 579396), which shows every node wired as intended.