GALIAIS-Nodes Scene Director

The name is a lie, and it's a good lie. The Scene Director sits in this pack's ai category, but it never calls an API, needs no key, and works fully offline. What it actually does is the clever part: it takes your tag prompt and rebuilds it as layered, natural-language scene directions - foreground, midground, background, lighting, atmosphere, camera - the kind of prose a natural-language model wants. It's a rule engine wearing a director's hat, and it's genuinely useful for exactly one job: converting your Danbooru-tag prompt into readable scene prose.

What you set

• 正面提示词 - the tag prompt to direct. Wired from upstream.
• 输出语言 - 中文 / 英文 / 中英混合, default 英文. This controls the language of the scene sentences it writes.
• 细节强度 - 精炼 / 标准 / 详细, default 详细. Verbosity of the output.
• 启用冲突剔除 and 冲突剔除策略 - before writing, it can prune conflicting tags (keep the earlier / keep the later / auto). Defaults: on, auto.
• 允许NSFW - off by default; in SFW mode it suppresses explicit rating tags.

How it works

The pipeline is deterministic and local. It parses your prompt, optionally prunes conflicts, then buckets each tag using the Danbooru taxonomy: pose tags become foreground direction, scene/object tags become midground and background props, lighting keywords (light, glow, shadow) get their own layer, atmosphere keywords (fog, mist, steam) another. Then it renders the buckets into structured English or Chinese sentences. No model, no inference - just taxonomy lookup plus sentence templates.

Three outputs come back: 场景导演提示词 (the natural-language scene text), 场景蓝图JSON (the structured foreground/midground/background breakdown), and 提示词段文本. Heads up: outputs one and three are the same string - the code duplicates it, so don't wire both expecting different content.

Where it fits

Reach for this when your target model wants prose instead of tags - the LLM-encoded models where a flat tag list is the wrong interface. It's also handy for img2img or multi-stage flows where you want the scene description as a human-readable artifact you can edit. The 场景蓝图JSON is the underrated output: it's the machine-readable scene plan, and you could feed it into downstream automation that this pack's AI nodes might consume later.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/GALIAIS/GALIAIS-Nodes.git

Restart ComfyUI. No dependencies. The optional DB input sharpens the tag-bucketing - without a loaded Danbooru DB, the taxonomy lookups fall back to keyword matching, so scene and prop tags get classified less precisely. It still works; it just gets dumber without the dictionary. And again: it will never ask you for a URL or API key, because it has no idea what those are.