GALIAIS-Nodes Danbooru Style Select

Style is the layer of an anime prompt that people iterate on most and organize least. The Danbooru Style Select is GALIAIS-Nodes' attempt to fix that: it gives you seven named style slots - rendering style, medium, color, lighting, post-processing, design style, quality detail - each with its own on/off toggle, so you build one style node and reuse it across every character instead of hand-tweaking a wall of tags.

It's the biggest node in the pack, and honestly it's two nodes in one: a tidy manual style composer, and a randomization engine for generating variations. Start with the first half and graduate to the second.

The manual half - what a beginner actually sets

Seven fields (all Chinese labels, as with the whole pack):

• 渲染风格 (rendering style), 媒介 (medium), 色彩 (color), 光照 (lighting), 后期效果 (post-processing), 设计风格 (design style), 质量细节 (quality detail).

Each field has a matching 使用… (use this field) toggle right next to it, and there's a free 追加到提示词 (append to prompt) field for anything that doesn't fit a slot. The 插入位置 dropdown (append / prepend / only) decides whether this block goes at the end, the front, or replaces everything else. 去重 (dedupe, on by default) stops duplicate tags leaking through.

Outputs are two: 提示词 (the joined style block - wire this into your composer) and 元信息JSON (what the node assembled). That's genuinely all you need for day one.

The random half - where the power is

Flip 启用随机Tag on and the node can fill empty (or every) field from the dictionary, keeping manual and random selections separate per the pack's design. There's a 随机策略 (only fill empty fields vs. append to fields), 每字段随机数 (count per field, with per-field overrides), 随机种子 for reproducibility, plus NSFW and heat floors. It re-runs on seed change, so it's a variation machine: same characters, different style each run, deterministic when you want it to be.

Beyond that there's a full AI-assisted mode - Tag生成模式 with options like "AI协同选择" and "AI意图定向选择" that need an AI服务商 wired in and use AI意图方向 (a rough description of the direction you want) plus RAG modes. The tooltip is honest about the constraint: the AI only picks from the enabled fields' candidate tags, and it writes the direction as natural language. Useful, but it's the advanced tier - get the manual and random layers working first.

Installing it

It's part of the GALIAIS-Nodes pack:

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

Restart ComfyUI, or grab it via ComfyUI Manager ("GALIAIS-Nodes"). No Python dependencies. It needs a dictionary (DB路径 or wired DB from the Danbooru DB Loader) for random fills, and an AI服务商 only if you're using the AI modes. The fields' tooltips on per-field random counts are worth reading in the UI - the -1 means "inherit the per-field default," 0 means "no random for this field."

The gotcha

The trap here is scope creep. With seven fields, eight toggles, random counts, AI modes and RAG knobs, it's easy to configure yourself into confusion. Keep it boring at first: fill the slots, leave random and AI off. Add randomization per field when you trust the output, and only then wire an AI provider. Style consistency is the point - if the node makes every run wildly different, you've turned on more layers than you meant to.