GALIAIS-Nodes AI Conflict Resolver

Danbooru tag prompts quietly contradict each other all the time. standing and lying, closed eyes and looking at viewer, blue hair and black hair - pick one, the model has to guess, and the guess is usually ugly. The AI Conflict Resolver is this pack's cleanup crew: you give it a prompt plus a diagnostics report, it asks your language model to spot the contradictions and hand back corrected Danbooru tags only. The whole point is a narrow, well-scoped job, which is exactly how you get an LLM to behave in a pipeline.

The three inputs

• AI服务商 - the provider object from GaliaisNodesAIProvider. Required, because this node makes a live API call on every run. Same "AI服务商" = AI service provider label as the rest of the pack.
• 提示词 (prompt) - the tag string you want audited.
• 诊断JSON (diagnostics JSON) - a JSON blob describing the conflicts. This is designed to plug in from upstream diagnostic nodes in the pack's Prompt Orchestrator / quality-check path, but you can paste a hand-written JSON object too. Feed it empty and the resolver still works, but it's sharper with a real diagnostics report.

Outputs are 修正建议 (the corrected/recommended tag list) and 原始响应JSON (raw API response for the audit trail). The 修正建议 string is what you'd splice back into your prompt or feed the next composition stage.

Where it earns its place

The reference workflow in the pack's README runs conflict pruning as an explicit stage of the Prompt Orchestrator. If you're building long multi-section character prompts with the 01–10 character nodes, the chances of contradictory tags climbing in across sections is genuinely high - one node says white hair, another says brown hair, and the composer happily joins them because it's deduping exact strings, not semantics. The resolver is the semantic net that dedup can't be.

The honest caveat is the same one for every node in this pack's AI family: it costs API tokens and needs a working provider. And it's an LLM giving opinions about tags, so treat its corrections as suggestions, not gospel - a confused model can strip a tag you actually wanted. Read 原始响应JSON if a fix looks wrong. For that reason I'd wire it in where you can review its output rather than blindly auto-merging.

Install and gotchas

Pack-wide install, once:

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

Restart ComfyUI (or ComfyUI Manager → search "GALIAIS-Nodes"). Zero pip dependencies, no bundled models. The pack's AI nodes all share two gotchas: you must configure the AI Provider first (run the Health Check before blaming this node), and the pack is brand new - no community signal online yet, so the JSON outputs and the README flow are the documentation. Start with a short test prompt before you point the resolver at a 200-tag character build.