XB-BOX - 🖼️ 生图提示词预设Pro

The menu name is Chinese - XB-BOX - 🖼️ 生图提示词预设Pro - which is why you probably googled the class name instead. It's three jobs in one node: it assembles your prompt, builds a correct empty latent for whichever 2026 model you're running, and can optionally hand the whole thing to a local or remote LLM for enhancement or image reverse-description.

2026 prompting has two traps: every model has its own dialect, and every model has its own geometry. Z-Image, Flux 2 Klein, Anima and Krea 2 are LLM-encoded, so bracket weights are literal punctuation. And the latent your sampler wants is 4-channel on SDXL, 16 on most of the new crowd, 128 on Flux 2, 64 on Hunyuan.

The preset half (works with the LLM switched off)

latent_kind is the important one: pick the model you're actually generating with and the node uses that model's official latent format - channels, downsampling divisor, minimum step, batch ceiling - then builds a real empty latent. In the pack's own source the table is Z-image 16ch/8/step 16, Flux2 128ch/16/16, Qwen-image 16ch/8/32, Krea2 16ch/8/16, Anima 16ch/8/8, Boogu 16ch/8/8, SDXL 4ch/8/8, SD3 16ch/8/16, Hunyuan 64ch/32/32. Pick wrong and ComfyUI's fix_empty_latent_channels still rescues you by adjusting channels and scale - you get a picture, not an error, just with a pointless conversion in the middle.

aspect_ratio, width, height and batch_size are the usual geometry controls, but with the grid step taken from the latent type rather than a hard-coded 16.

preset_mode and three_view_text are the interesting pair. On 常规文生图 (plain text-to-image) your body text passes straight through. The three-, four- and five-view character modes and 背景纯透明 prepend a preset sentence describing a turnaround sheet - front/side/back panels plus a face close-up - or an RGBA transparent-background image. Edits to it are stored per mode and per language, so switching modes doesn't wipe them - the character-sheet pattern the consistency crowd has run for years, now a dropdown.

output_lang (中文/英文) governs the preset sentence, the element-panel separators and, with the LLM on, the language you get back. The panel itself is an 8-category chip picker in the frontend - style, camera, subject, pose, outfit, props, lighting, background.

The LLM half - and no, it isn't lying when it's off

use_llm defaults to false, and in the source both the model load and the API call sit inside that branch. Off means off - no GGUF, no request, no key. Turn it on and the node builds a system prompt from preset (38 model-specific enhancer presets in EN and ZH - Z-Image Turbo, Flux.2 Klein, Qwen-Image, Krea, Wan), your language, and task_preset (23 reverse/instruction presets, where * is a required placeholder and # is where your text lands), then demands only the final prompt back.

The task changes with what you feed it. Text only: enhance. Image only: caption. Image plus text: the text becomes an edit instruction, and the node forbids "originally X, now Y" comparison prose - it spots that phrasing and retries once. Anyone who has chained an edit model off a VLM knows why.

backend picks the engine: 本地模型 runs llama-cpp-python on your own GGUF; 在线 API uses whatever the config dialog saved (OpenAI-compatible - DeepSeek, Qwen, GLM, Kimi, Ollama, vLLM, LM Studio - plus Anthropic and Gemini). Keys live in ComfyUI's user directory, not the workflow: great for sharing, plaintext on disk, so treat them as secrets.

Outputs: 提示词 (the final prompt → CLIP Text Encode), 空latent (→ your sampler), 提示词列表 (one line per entry, for list consumers), 提示词设定 (the system prompt that ran - your debugging window). Optional inputs: text (external prompt; wired in, it wins and greys out the node's own box) and images for the vision path.

Install

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

Or Manager → search XB_ToolBox. The preset half needs nothing beyond torch and ComfyUI itself.

For the local half you install llama-cpp-python yourself - it's deliberately commented out of the pack's requirements.txt. Nvidia users take a prebuilt CUDA wheel matching their Python version from the JamePeng releases; ROCm users compile. Put the GGUF (and, for VLM reverse-description, its matching mmproj) in ComfyUI/models/LLM/. The config's vram_limit then works out how many layers fit on your card.

Where people get burned

"未选择本地模型" on first run. The error itself tells you to open the 🤖 model-config dialog and pick a model. That dialog state lives in the node's internal settings JSON, so a half-configured node travels with your workflow.

Trusting the pack README over the pack. "No extra pip dependencies, plug and play" isn't true of the whole toolbox anymore, and neither prompt/LLM node appears in the README at all - everything above comes from the shipped source. It's also an LLM node, so it's expected to touch the network and load weights: skim a fresh pack before its first run.