GALIAIS-Nodes 07 叙事主题

Most prompt builders stop at "who, what, wearing what." The Narrative node (07 叙事主题) is where this pack gets opinionated: it adds the story layer - what's happening between characters, what mood the image carries, what reference or meme it's nodding to. On a model trained on Danbooru, where every image has an implied story in its tags, this layer is how you tell the sampler which story you meant.

What you fill in

Three content fields plus a supplement, each with a 使用 toggle:

• 关系事件 (relationships/events) - the relational story: lovers, rivals, birthday, confession, farewell, whatever the scene is about.
• 主题状态 (theme/state) - the atmosphere or condition: loneliness, victory, melancholy, drowsy, celebration.
• 引用梗 (reference/meme) - the nods: JoJo pose, zombie hand, screaming face, any recognizable reference or meme tag.
• 叙事补充 (narrative supplement) - free-text extras, multiline.

Plus the shared layer controls: 角色槽位 (全局 / 角色1–8), 角色标签, 启用, 权重 (0–5), and the random/AI block (启用随机Tag, Tag生成模式, 随机种子, optional AI服务商), with optional DB, 上游上下文, and 上游提示词段. For scene-wide narrative, 全局 is the natural slot.

Outputs

The standard trio: 叙事段 (section object → composer), 文本 (plain text), 元信息JSON (contribution + random audit).

Where it earns its keep

The 关系事件 and 主题状态 fields are where "two characters in a room" becomes a scene with tension or warmth. Combined with the 02 layer's 情绪表情, they're the difference between an image and a moment. The 引用梗 field is the pack's quiet crowd-pleaser - meme and reference tags are a well-known weak spot for LLM rewrite tools (they get "cleaned up" into blandness), so having a dedicated, un-touched field for them is genuinely thoughtful.

The honest read

This is a soft layer. Its tags don't shape pixels the way 镜头构图 or 服装 do - they shape mood and semantics, which means their effect ranges from strong to invisible depending on your base model's training. On tag-hungry anime-SDXL models (Illustrious, NoobAI, Pony) the relationship tags read well; on natural-language checkpoints they mostly get ignored, so don't be shy about writing short prose here for those. Also: 主题状态 tags overlap with the 02 layer's expression tags, so don't duplicate - pick one layer as the authority per mood. The composer dedupes exact strings only, so melancholy in both layers will render twice.

Install

Pack-wide, once:

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

Restart ComfyUI, or ComfyUI Manager → "GALIAIS-Nodes". No pip dependencies; the Danbooru DB is separate if you want dictionary-backed picking. Young pack, no community writeups - the 元信息JSON and README are the docs.