GALIAIS-Nodes Prompt Orchestrator

This is the pack's flagship: the node that turns the whole recommended flow into a single step. You feed it a rough positive prompt, it prunes contradictory tags, builds a structured blueprint, optionally hands it to an LLM for enrichment, assembles the negative, and returns the final positive, negative, natural-language draft, and a quality score - all in one run. The README's step-eight "orchestrator" is exactly this node.

It's the node you build toward. When you're tired of wiring the blueprint, enricher, and inspector together by hand, this collapses the whole pipeline.

How it works

Required inputs: AI服务商 (the provider object - this node genuinely needs one for its optional AI step), 正面提示词 (positive prompt), 负面追加 (extra negatives), and 负面预设 (negative preset, same list as the Negative Preset node). Then the pipeline controls:

• 启用冲突剔除 with 冲突剔除策略 - deterministic conflict pruning runs first, so contradictions never reach the LLM.
• 启用AI扩写 (AI expansion, on by default) with AI扩写模式 (the same five modes as the enricher: natural-language supplement, Anima captions, scene director, tag-constrained expansion) and 细节强度.
• 输出语言, 允许NSFW, 去重, and 失败时返回原文 (on by default here - the pipeline fails softly, returning your original prompt rather than erroring out).

Internally it chains exactly the nodes you'd wire yourself: the Image Detail Blueprint (deterministic, no API) produces the structured plan and the scene/full drafts; then, if AI expansion is on, the Positive Prompt AI Enricher expands the blueprint prompt (the constrained, conflict-free version - the pack's answer to subject drift); the negative is assembled from preset + your 负面追加; and the final prompt runs through the same diagnostic that powers Prompt Inspector V2.

Five outputs: 最终正面提示词, 最终负面提示词, 自然语言, 质量评分 (the diagnostic score), and 流程JSON (the full audit trail - which AI mode ran, what was pruned, the raw LLM response, the quality report).

Installing it

It ships in the GALIAIS-Nodes pack:

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

Restart ComfyUI, or install via ComfyUI Manager ("GALIAIS-Nodes"). No Python dependencies. You need an OpenAI-compatible provider and a key (literal, or env:VAR / $VAR), though the deterministic blueprint stage runs even without it if you disable AI expansion. Field labels are Chinese, as throughout the pack.

The honest verdict

This is the node that makes the pack feel "enterprise," and it's genuinely convenient - one run, final prompts out, audit trail in JSON. The tradeoffs are the same as any LLM-in-the-loop node: AI expansion adds latency, costs per call (unless local), and needs the leash kept short. If you run it and the output wanders, trim the sentence budget (the enricher's 最少/最多句数) or drop AI expansion and keep the deterministic pipeline. And because responses are cached, a "stuck" output is often the cache (see DB Cache Control). When it's tuned, it's the closest thing this pack has to "push button, get prompt."