XB-llama - 🎞️ 分镜词处理器

If you've ever run an LLM storyboard workflow you know the problem: the model hands you a single wall of text with a dozen shots in it, and you want to generate one of those shots without re-encoding the whole thing. XB_llamaStoryboardProcessor is the little slot machine in the middle of the XB-llama chain that splits that wall of text into cards and lets you draw from the deck.

It's part of the XB_ToolBox "XB-llama" family, which is a llama-cpp-python reimplementation of the ComfyUI-llama-cpp_vlm nodes with extra AMD/ROCm handling. The processor sits downstream of the storyboard writer - feed it the multi-line output where every shot starts with 分镜图_01, 分镜图_02, or Scene_1 - and it caches those lines, then hands you a single list you can wire into your text encoder.

How it works

The mechanism is simpler than the Chinese UI suggests. It regex-matches each line for a 分镜图_\d+ or Scene_\d+ header and stashes the whole text in a class-level cache. Then it has two personalities:

• draw_mode off (default): it's a pass-through. It strips blank lines and emits every line as a list item, so you get a clean per-shot list.
• draw_mode on: this is the "抽卡" (draw/gacha) mode. It ignores new upstream input and serves from the cache instead, letting draw_range pick which shot to output - 全部 for everything, or a single 分镜图_01 through 分镜图_16.

That cache trick is the whole point. Because the upstream LLM gets skipped in draw mode, you can generate shot 7 twelve times, tweaking the seed, without the language model ever running again. Wire the processor's output to your CLIP/text encoder (or straight into a video i2v node that takes a prompt), and each shot prompt becomes its own generation.

The inputs that matter

Only three, and only two of them will ever bite you:

• draw_mode (BOOLEAN) - the off/on switch for gacha mode.
• draw_range (enum: 全部 + 分镜图_01…分镜图_16) - which shot to draw.
• text_input (optional, any type) - the storyboard text from upstream. Marked optional because the whole design assumes it can go dark while the cache serves.

Output is output_list, a list of STRINGs. In normal mode each line is one item; in draw mode it's the shot (or shots) you picked. If draw mode can't find the requested shot in the cache, it doesn't error - it emits a placeholder prompt that just renders a solid-color image with "未找到分镜图_XX" text on it. Confusing the first time you see it, actually helpful once you realize it's telling you the header name doesn't exist in your text.

Install

The pack itself:

cd ComfyUI/custom_nodes
git clone https://github.com/WJLUOXIAO/XB_ToolBox.git
# or: ComfyUI Manager → search "XB_ToolBox" → install

The README swears there are zero pip dependencies, and for the core pack that's mostly true. But the whole XB-llama family only loads when llama-cpp-python is installed - without it these nodes don't appear in your node list at all. Install it built for your backend:

pip install llama-cpp-python

On an AMD card you want the ROCm/HIP (or Vulkan fallback for older cards) build; the nodes auto-detect which GPU family they're on.

Common issues

• Node missing entirely - llama-cpp-python isn't installed, or the module import failed. Check your ComfyUI console for a line like [XB-llama] llama-cpp-python 未安装.
• Draw mode says "暂无缓存" (no cache) - you ran it with draw mode on before ever running it in pass-through mode. Run once with draw_mode off so it can cache, then flip the switch.
• The "未找到" placeholder image - your cached text uses a different header name than the range you picked (e.g. Scene_1 vs 分镜图_01). The regex accepts both, but the dropdown only knows 分镜图_01-style names, so check which one your writer actually emitted.