🎞️ Scene

Scene is the assembler. Every other node in the ComfyUI Scene Composer pack builds one slice of a prompt - the shot, the pose, the character, the outfit - and Scene is the node that collects those slices and staples them into a single finished prompt string you can hand to CLIP Text Encode. If the pack is a factory line, this is the end of the belt.

That matters because the rest of the pack is deliberately fragmented. Scene Composer, by taches, procedurally generates NSFW anime scenes by having each concern live in its own node, so you can reroll the outfit without touching the pose, or lock the camera and spin the action. But eventually all those loose strings need to become one prompt in the right order - and that's the only job Scene has.

How it works

Scene takes up to five text fragments, concatenates them in a sensible scene order, and emits one combined STRING. No model, no inference - it's string plumbing with a seed on top. Everything upstream already did the creative work of picking tags; Scene just decides the final layout and hands you the result.

The tag order it produces is the point. On Pony and Illustrious-family models, earlier tags get stronger attention, so the composition/quality preamble wants to come first and the finer details later. Scene bakes that ordering in so you don't have to think about it.

The inputs

There's exactly one required input and five optional ones:

• seed (required) - the reproducibility knob. Fix it and the assembled prompt is stable run to run; change it to shuffle anything the upstream nodes left on random.
• composition (optional STRING) - from the Composition node: the quality preamble, how many people, the camera framing and angle. This is your "first tags," so it belongs here.
• action (optional STRING) - from the Action node: pose, gesture, and (if enabled) the explicit act.
• character (optional STRING) - from the Character node: the randomly generated person - hair, eyes, build, and so on.
• clothes (optional STRING) - from the Clothes node: the outfit.
• environment (optional STRING) - the setting. Note there's no dedicated environment generator among these five nodes, so you feed this one a plain string yourself: a simple text/primitive node with something like bedroom, indoors, night, or whatever the pack's own environment node produces if you've got it.

They're all optional, which is genuinely useful: wire in only the pieces you want procedurally generated and leave the rest empty. Want a fixed character in random poses? Feed character a static string and let Action roll. The single output STRING goes into the text input of your positive CLIP Text Encode (Prompt) node, and from there it's a normal generation - encode, sample, decode.

The context you need: this is a Pony rig

The pack's default quality string is score_9, score_8_up, score_7_up, source_anime - that's Pony Diffusion V6 XL's score-tag opener, so Scene's output is built to be read by an SDXL anime checkpoint (Pony, or the Illustrious/NoobAI family with the quality tags swapped). Booru tags like these do real work on those models and roughly nothing on Flux or the 2026 LLM-encoded models, which expect sentences. So before you blame the node for a bad image, confirm the checkpoint on the other end actually speaks this language.

Installing it

Scene ships with the whole pack:

• ComfyUI Manager - search ComfyUI Scene Composer and install, or import the repo's workflow.json / workflow.png and click Install Missing Nodes. Restart after.
• 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.

No heavy model files come with the pack - these are text nodes - but you do need a Pony or Illustrious checkpoint to render. And heed the README's warning: v0.3.x moved to ComfyUI's subgraph system and old saved workflows are expected to break, so build from the repo's current workflow.json.

Common issues

If the prompt comes out with pieces missing, check that the upstream node outputs are actually wired into the matching Scene inputs - every one is optional, so an unconnected slot silently contributes nothing rather than erroring. If the whole thing renders as generic anime with none of your intended scene, you're likely on the wrong model family. And since this is a small, one-author project with almost no community discussion to lean on, the canonical example of Scene wired correctly is the author's full NSFW workflow on CivitAI (model 579396) - when in doubt, copy that graph rather than guessing.