镜头分词器 / Shot Splitter

If your workflow involves turning a written storyboard into actual frames, you've probably felt the pain this node removes. You ask an LLM to write a scene-by-scene script, it hands you one giant blob of text, and now you need to break that blob into individual shot descriptions before you can generate one image per shot. Shot Splitter does exactly that: it takes a multi-shot script and outputs a proper list, one entry per shot, ready to feed a batch-generation loop.

It's one of the text-tooling nodes in RUI-Nodes - the pack's author clearly builds comic and animation pipelines, because this node plus Dialogue Extractor and Page Narration Remover form a little storyboard-to-render assembly line. The workflow shape is: LLM writes the script → Shot Splitter breaks it up → the list feeds an image-generation batch → Dialogue Extractor pulls the voiceover lines. If you're doing anything like that, these three nodes are why the pack is worth installing.

How it works

The input text is expected to wrap each shot in explicit tags:

<SHOT_1>
the first shot's description
</SHOT_1>
<SHOT_2>
the second shot's description
</SHOT_2>

The node splits on those tags and returns shot_descriptions as a LIST - one string per shot, in order. Then two optional knobs let you pull a slice instead of the whole thing: start_shot_num (0 = start at the first shot) and shot_count (0 = take everything through the end). Want shots 3 through 5? start_shot_num = 3, shot_count = 3. The summary output is a plain string with an overview of what was extracted - handy for logging or feeding a display node.

The inputs you'll set:

• input_text - the multi-line script, with <SHOT_XXX> tags.
• start_shot_num / shot_count - optional range slicing.

Outputs: shot_descriptions (LIST) and summary (STRING).

Installing it

Part of the RUI-Nodes pack ("Rui-Node🐶"). Install once via ComfyUI Manager (search "RUI-Nodes") or:

cd ComfyUI/custom_nodes
git clone https://github.com/rui40000/RUI-Nodes
cd RUI-Nodes
pip install -r requirements.txt

Restart after. Pure Python text processing - no extra dependencies.

Where people get burned

The format requirement is the whole game. If your LLM outputs shots with markdown headers or arbitrary separators instead of the <SHOT_N> tags, the node has nothing to split on and you'll get one giant entry. The fix is in the prompt: tell the model to emit the tags. Also worth noting the node recognizes a few spellings of the tag ("镜头1", "分镜2", "Shot 3" are mentioned in the tooltip), so it's more forgiving than strict XML - but your best bet is still to make the source script well-formed. And since shot_descriptions is a list, it feeds list-consuming batch nodes directly; if your target wants single strings, wire in a list-to-string step.