ZML_随机文本权重

ZML_随机文本权重 is the "surprise me" node: it picks random lines or tags from a text file and slaps a random weight on each one, then outputs the result as a prompt fragment. Out of the box it ships with a 1000-artist danbooru tag file - the kind of thing you point at when you want a random artist style, or a random tag salad for testing - and you can point it at any txt file you drop into zml_w/txt.

The mechanism is simple and worth knowing because of one detail: the node uses random.sample (no repeats) to pick 随机个数 items, and each item gets a weight drawn uniformly between your 最小权重 and 最大权重, formatted to the precision you chose. Trailing commas get stripped so the output is clean for joining. And the always-dirty trick is in play - its IS_CHANGED returns NaN, which the KB's plumbing essay names as the canonical way to force a node to re-run every time. In plain terms: it re-rolls on every execution, no cache fights.

What to set

• 文件 - dropdown of txt files in zml_w/txt, or 禁用 (使用文本框) to use the optional text box instead.
• 随机模式 - 按行随机 (random whole lines) vs 按标签随机 (split on commas and random tags).
• 小数位数 - 两位 or 一位 decimals on the weights.
• 随机个数 - how many lines/tags to pick.
• 最小权重 / 最大权重 - FLOAT 0–3, default 0.3/0.5. The bounds for the random weight.
• 文本 (optional) - your own tags/lines when the file selector is disabled.

Outputs: 文本 (the weighted fragment) and help (the author's own usage instructions - a rare, genuinely useful extra; wire it to a text display).

The honest take on weights

Here's the part that requires judgment, and the KB's prompt-engineering refresh is the grounding. Random tags are a time-honored way to discover new directions in SDXL-lineage models (Illustrious, NoobAI, Pony, Wan) where the encoder is CLIP and (tag:1.2)-style weights work. On the 2026 LLM-encoded models, attention weights are silently discarded and tag lists aren't how the prompt is read - so this node is great fun on the SDXL family and close to meaningless on a Z-Image/Flux-2-style encoder. Know which one you're on. The author builds this kind of thing for their Wan/SDXL rotation, which tells you its natural habitat.

The other thing to know: it's a one-person, Chinese-first pack (translation patch in the README), and the node is thin by design - pick, weight, output. It's a discovery tool, not a prompt engineer. Pair it with the pack's text-format node if you want the weighted output normalized into your SDXL/NAI weight syntax.

Install

ComfyUI Manager → search ComfyUI-ZML-Image, or:

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

restart. The artist txt files ship with the pack - no downloads needed. Drop your own .txt into zml_w/txt (top level) and it appears in the dropdown after a restart.