GALIAIS-Nodes AI Negative Prompt Builder

The negative prompt is alive and well on the anime SDXL lineage - Illustrious, NoobAI, Pony - and this node exists to stop you from writing the same bad hands, extra fingers, mutated hands list from memory for the hundredth time. You give it your positive prompt and pick a preset, and it has your LLM build a comma-separated negative list tuned to what's actually in the positive. Then it merges your chosen preset in on top, deduplicated.

That last bit is the design worth noticing: the output is the preset plus the model's additions, joined cleanly. So you get a solid baseline negative even if the model under-delivers, and the model's contribution is scoped to the one thing you asked - "build a negative prompt for this positive prompt, return comma-separated negative tags."

The inputs that matter

• AI服务商 - the provider object. Live API call, so required, same "AI服务商" label as the rest of the pack.
• 正面提示词 (positive prompt) - your positive. The model uses it to decide what to negate, so "write a bad-hands negative" makes a lot more sense when it can see the pose tags.
• 负面预设 (negative preset) - this is the genuinely useful dropdown, and it's pure rule-based, no API needed:
  • 标准 (standard) - the classic worst quality, low quality, bad anatomy, bad hands... block.
  • 轻量 (light) - minimal: worst quality, low quality, blurry, watermark, text.
  • 手部修复 (hand repair) - pure hand-focused terms, for when hands are the only problem.
  • 写实 (photoreal) - exposure/quality terms tuned for photorealism.
  • 无 (none) - no baseline, model-only output.

Outputs: 负面提示词 (the final merged negative) and 原始响应JSON (raw model response).

The honest take

The preset dropdown is the star; the AI half is nice-to-have. Honestly, for most anime-SDXL work the 标准 or 手部修复 preset alone is 90% of what you'd get from the model, and you could skip the API call entirely by wiring the preset into the composer. Where the AI adds value: your positive has unusual failure modes (a specific bad-prop tendency, a specific anatomy quirk) and you want the model to name the tags for them.

Two notes. First, this node's premise assumes your checkpoint still uses a negative prompt - on 2026 flow-matching models with distilled guidance and an LLM encoder, the negative does almost nothing and you should skip this entirely. Second, it's an API node: tokens per run, provider required, and if the model starts listing stuff that isn't in your scene, the preset merge keeps you safe. Read the 原始响应JSON before trusting the model's additions.

Install and gotchas

Pack-wide install:

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

Restart ComfyUI, or find it via ComfyUI Manager ("GALIAIS-Nodes"). No pip dependencies. The usual pack gotchas apply: this is a brand-new, low-signal pack (no community writeups), and if you hit API errors, configure the AI Provider and run its Health Check before debugging the writer.