闲兔｜数字人提示词输出

One input, one output, one job: it reads the digital-human workbench state and returns the current segment's prompt as a STRING. The reason it exists is that the workbench's prompts are trapped in a panel. You typed them there, the panel updates them as you move between segments, and without this node there's no socket to send them anywhere.

The important nuance, because it trips people up: this returns the prompt you wrote, not the rewritten one. The optimized version is what the Qwen H3 Prompt node produces, and it goes into the studio's optimized_prompt input or the segment's stored optimized_prompt. If you want the LLM's version elsewhere in the graph, take it from ComfyUI_Qwen_H3_Prompt's own output. This node is the raw text.

Why that's useful

Local prompt enhancers are a rewrite you didn't fully control - that's the standing lesson from the LLM-in-the-graph literature, where the two failure modes are dirty chat scaffolding leaking into your conditioning and the enhancer inventing detail you never asked for. Keeping a socket on your original text lets you A/B it: same reference images, same sampler, your sentence against the model's, and you can actually see whether the optimizer helped.

The other real uses are mundane and good. Route it into a CLIP Text Encode or a Qwen Image text encoder to drive a completely different graph from the same storyboard. Pipe it into a Show Text / save-to-file node so you have a written log of every shot prompt next to the renders. Or feed it to a text-to-video model that speaks natural language - H3's own prompt handling, LTX-2, Wan - where the digital-human panel was just a convenient place to write twelve prompts in order.

Mechanics and errors

studio_state is the workbench JSON; the node parses it with the pack's shared reader, which validates the JSON, requires at least one internal segment and caps you at 500. It then resolves the active segment (the one highlighted in the panel) and returns prompt verbatim - no prefixing, no negative prompt merge, no camera text appended. If the active segment's prompt is empty you get 当前分镜的提示词为空。, which is the correct behaviour but does mean an empty segment can't silently pass through as a blank string.

Like its sibling output nodes, it is not an output node: running it writes nothing. If it's not wired to something, it does nothing at all.

Install

Included with the pack - no separate install:

cd ComfyUI/custom_nodes
git clone https://github.com/xiantuhenlan/ComfyUI-Xiantu-QwenImage21-Workbench.git
pip install -r ComfyUI-Xiantu-QwenImage21-Workbench/requirements.txt

Restart ComfyUI and force-refresh the browser afterwards; the pack's workbench panels are served from a versioned frontend directory, so a hard refresh is the difference between "working" and "where did the panel go". As always with this pack: no models, no workflows, and nothing else to download for this node specifically.