2🐕Scene class

The README calls this whole family of nodes the "Hint Word Master" - a set of curated, dropdown-driven prompt builders you use instead of typing tag lists from memory. Scene class is the one that covers where your image is happening: background type, sky, indoor and outdoor location, building, and overall mood, each picked from a long menu instead of typed freehand, each with its own attention weight.

This matters more on SDXL-lineage checkpoints (Illustrious, Pony, and their relatives) than on the newest LLM-encoded models, where the KB's own prompt-engineering notes are blunt about it: (word:1.3)-style attention weights get silently discarded on models with an LLM text encoder wrapper. If you're on one of those newer models, the weight sliders here won't do anything - the categories and tags themselves still work fine as plain text, you just lose the emphasis control. If you're on SDXL or an anime-tuned checkpoint, the weighting is very much alive and this node leans into it.

The categories

Seven pickers, each paired with its own weight slider (0.1-2.0, default 1): Background (simple/white/gradient/transparent background, seasonal settings like Halloween or Valentine's, sci-fi, and more - 16 choices total), Sky (day, night, clouds, stars, sunset, rain, and more - 27 choices), Indoor (classroom, bedroom, kitchen, library, gym, and dozens more - 34 choices), Outdoor (city, road, bridge, alley, graveyard, and many more - 86 choices), Building (onsen, ruins, castle, skyscraper, shrine, rooftop, and more - 85 choices), and Scene Atmosphere (dark, mystical, whimsical, tranquil, intense, and more mood descriptors - 29 choices). Any of them can be set to None to skip that category entirely.

There's an eighth one worth calling out on its own: Architect, 71 choices - and despite the name, the sample list is entirely film directors (Wes Anderson, Bong Joon Ho, Charlie Chaplin, the Coen Brothers, and so on). That's almost certainly a rough English rendering of a Chinese label that meant something closer to "director," not an actual architecture category - worth knowing so you don't skip past it looking for building styles and miss what it actually does, which is nudge the whole image toward a particular filmmaker's visual sensibility.

random (yes/no) and seed work together: set random to yes and the node picks entries for you instead of using whatever's selected in each dropdown, reseeded by seed. The single output is prompt, a composed STRING ready to feed into your positive prompt (typically concatenated with your subject description via one of this pack's text-splicing nodes).

Installing it

ComfyUI Manager: search "Comfyui-ergouzi-Nodes", install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/11dogzi/Comfyui-ergouzi-Nodes.git

No models needed - every choice here is a plain text tag baked into the node, nothing downloads. Standard note for the pack: this English repo is the frozen branch, with active development on a separate Chinese-first sibling.

Common issues & troubleshooting

Weights don't seem to do anything. Check your checkpoint family first - if you're on a model with an LLM text encoder, attention-weight syntax is discarded outright, and that's a model-level behavior, not something this node can override.

Prompt gets too long / cluttered. With eight weighted categories in play at once, it's easy to end up with a dense prompt block that fights your subject for attention. Set the categories you don't care about back to None rather than leaving every slot populated by default.