XB-BOX - 🖼️ Qwen2.1提示词预设

The "text encoder" on modern image models isn't a little CLIP sidecar any more. On the Qwen-Image line it's a full vision-language model, and when ComfyUI loads it as CLIP it still holds the whole thing in VRAM. It can write.

This node exploits that. Click a few style/view/subject chips, type "a girl in a coat on a pier", hit run - and the encoder rewrites it into a long structured prompt, locally, offline, no API key, no Ollama, no second download. Mechanically it's ComfyUI's native Generate Text path, minus the assembly work.

What you're actually getting

Two halves in one node. The preset half is the same panel the pack's XB_ImagePromptPreset uses: 15 preset modes, split into a text-to-image group (三视图/四视图/五视图 character sheets, 背景纯透明, 图文版面, 信息图, 多格分镜, 广告分镜板) and an image-to-image group (保持主体换场景, 局部编辑, 老照片修复, 整图风格化, 360°全景, 多图指认合成) - plus an eight-category element palette (风格 / 视角 / 主体 / 姿态 / 装扮 / 道具 / 光影 / 背景).

The generator half is the official Load CLIP and Generate Text path, copied field-for-field.

The pairing is the point. prompt-engineering.md argues that on an LLM-encoded model your prompt is an instruction, not a token bag, so having a language model compose it is the right shape. The preset half keeps the composer honest: the output is the preset's 设定词 placed verbatim at the top, with the generated body underneath. That fixed header is your anchor against enhancer drift.

The mechanism, briefly

Your text goes in as 【正文】, the preset text as 【参考设定】; both are tokenized together with the encoder. Generation runs at your sampling settings, the result is decoded (a <think> block, if any, is stripped), and the node assembles 设定词 (untouched) + generated body with the right framing clauses. Two outputs:

• 提示词 (STRING) - wire this into your text encode node, or into an edit encoder on the image side.
• 空latent (LATENT) - wire this into the sampler. On the image-to-image path, don't: the official edit chain is input image → VAEEncode → sampler.

No negative prompt here, and that's correct: negatives are inert on LLM encoders.

The fields worth touching

Most inputs hide behind two popup buttons (🤖 LLM设置, ✨ 预设参数). The ones you'll actually set:

• clip_name / clip_type / device - the Load CLIP trio. Drop your text encoder in ComfyUI/models/text_encoders, pick it, leave clip_type on qwen_image, device on default. cpu trades speed for card space. A mismatched type is the classic load failure.
• preset_mode and io_mode - pick a mode; 自动 already checks whether an image is connected and picks the right framing (and the right SKILL file). Forget the reference image on a 图生图 mode and it logs a warning and falls back to plain text.
• latent_kind / aspect_ratio / width / height / batch_size - nine latent kinds (Anima, Boogu, Flux2, Hunyuan, Krea2, Qwen-image, SD3, SDXL, Z-image), each with its own official size step, so width snaps to a legal value.
• max_length - the 512 default is thin for 图文版面, 信息图, 多格分镜 and 广告分镜板; the author's own guidance is 1024–2048 or your layout rules get truncated mid-sentence.
• sampling_mode / temperature / seed - off is greedy decoding and ignores the sampler settings entirely. Want the same prompt twice? Use off, or fix the seed.

The optional text input overrides and locks the on-node prompt box; images takes a batch as N reference images (up to 10, referenced by number inside the prompt).

Installing it

ComfyUI Manager, search XB_ToolBox, or:

cd ComfyUI/custom_nodes
git clone https://github.com/WJLUOXIAO/XB_ToolBox.git

Restart, then point it at your Qwen-Image text encoder. No models ship with this node, and no key is needed.

One caveat: the README claims "NO extra pip dependencies required! Plug and play." The repo's requirements.txt disagrees (PyAV, easyocr, opencv, onnxruntime, pythonnet), and nodes_audio_slicer.py imports av at module top, pulled in unguarded by __init__.py. If the pack never appears in your node list on a bare environment, that's usually why:

pip install -r ComfyUI/custom_nodes/XB_ToolBox/requirements.txt

Where people get burned

There's no bypass. Execute it and it generates, every queue, even if all you wanted was the latent. On a slow card that's real seconds - minutes with the long output modes.

The encoder cache holds one entry keyed on name/type/device, so alternating between two text encoders reloads each time. Selecting a SKILL file does nothing unless skill_mode is 手动, and SKILL does nothing at all when use_default_template is off - same as the official node ignoring system_prompt. JSON output with a rewritten_prompt field is the bundled skill files' contract, not a break. And on AMD/APU boxes where the iGPU enumerates first, HIP error: device kernel image is invalid means ComfyUI landed on the wrong device - relaunch with HIP_VISIBLE_DEVICES=1.