ZML_文本转格式

ZML_文本转格式 is a prompt janitor. Paste in a messy tag list - NovelAI weight syntax, broken punctuation, duplicated tags, stray underscores - and it comes out the other end clean, SDXL-ready. It's the pack's "format this prompt before it hits the encoder" node, and if you trade prompts between A1111/NovelAI and ComfyUI it will save you a ton of hand-editing.

The headline feature is NAI → SD weight conversion. NovelAI writes emphasis as {tag} (boost) and [tag] (reduce), and both SD 1.5 and the SDXL-lineage anime checkpoints expect (tag:weight). The node walks your text, counts brace depth on the { and [ runs, and converts:

• 精确 (exact) mode: {tag} → (tag:1.1), {{tag}} → (tag:1.21), [tag] → (tag:0.9) - exponential, matching the NAI behavior
• 一位小数 (one decimal) mode: linear (tag:1.1) / (tag:1.2) style
• 清空权重 strips all weight syntax entirely, leaving bare tags

The 文本格式化 dropdown handles the cosmetic passes: underscores to spaces, spaces to underscores, comma-separated tags, comma-to-newline, stripping newlines, and 括号转义 (escaping literal parens). Then 格式化标点符号 (on by default) tidies commas, periods and BREAK into SDXL-safe form - the README describes it as fixing "non-conforming punctuation." Finally, 合并相同提示词 dedupes repeated tags, with 合并白名单 protecting a comma-separated list of phrases (default BREAK) from being merged away.

Single 文本 output, plain STRING. Wire it between your prompt source and a CLIP text encode, or in front of the pack's own filter node. If you're moving an A1111 or NAI prompt into ComfyUI, this is the natural place to normalize it - the KB's prompt-engineering doc is blunt that (word:1.3) style weights are alive on the SDXL anime lineage (Illustrious, NoobAI, Pony), which is exactly the audience this node is aimed at.

Installing it

Part of ComfyUI-ZML-Image. ComfyUI Manager → search ComfyUI-ZML-Image, or:

cd ComfyUI/custom_nodes
git clone https://github.com/zml-w/ComfyUI-ZML-Image

Restart. All the conversion logic is pure Python + regex, so no models or downloads - the node's heaviest dependency is Pillow, already in ComfyUI. Chinese UI by default; English patch at https://github.com/zml-w/ZZZ_ZML_English_Patch.

Common issues

The weight conversion is the part that can surprise you. It assumes the NAI {}/[] convention and converts those; if your prompt already uses SD (tag:1.2) syntax, leave 权重转换 off or you'll get nested parens that look fine and behave wrong. The 括号转义 option exists precisely because of that collision. Also, the exact-mode math is 1.1^n and 0.9^n - a triple brace becomes 1.331, which is hot. Most people keep NAI weights to single or double braces.

Dedupe is worth testing on your own prompts before trusting it - the whitelist default (BREAK) shows the author hit the same wall you will: naive dedupe happily merges "BREAK" segments you wanted kept apart. If you see tags disappearing you wanted, that's the 合并白名单 you need to fill in.

One honest note: the UI is Chinese, and the "帮助" help text was removed from this node in an update (the README says so). If a dropdown option isn't obvious, this guide covers the four that matter.