XB-BOX - Image Prompt Preset

What it replaces

Every text-to-image workflow opens with three separate things: a prompt box, a resolution, and an EmptyLatentImage for the sampler. Change the ratio and forget the latent and you get a stretched or silently resized image. XB_ImagePromptPreset is all three in one, from XB_ToolBox ("XB-BOX") - a Chinese-authored beginner toolbox, young enough (v0.9.01) that there's essentially no community footprint to judge it by. So judge the code, which is short.

The two outputs

提示词 / "Prompt" (STRING) and 空latent / "Empty Latent" (LATENT). String into your positive CLIP Text Encode, latent into your sampler. There's no negative output - fine on CFG-1 distilled models like Z-Image Turbo or Flux 2 Klein, where the negative pass isn't computed at all, and a reason to encode one separately on SDXL.

The latent dropdown is the interesting part

latent_kind covers nine families - Z-image (the default), Flux2, Qwen-image, Krea2, Anima, Boogu, SDXL, SD3, Hunyuan - and each carries a real spec, not a label: channel count, downsample divisor, size step. SDXL is 4-channel/÷8, step 8. Z-Image, Qwen-Image, Krea2 and SD3 are 16-channel/÷8, step 16. Anima and Boogu are 16-channel, step 8. Flux 2 is 128-channel/÷16. Hunyuan is 64-channel/÷32.

The mechanism: an empty latent is a zeros tensor shaped [batch, channels, height÷divisor, width÷divisor], plus one piece of metadata - downscale_ratio_spacial - that ComfyUI's comfy/sample.py reads to decide whether your latent needs rescaling to fit the model you actually load. Channel count isn't cosmetic: a latent from the wrong VAE family gives noise or flat colour, not a subtly wrong picture.

Because it's all zeros, though, a wrong pick usually won't crash - ComfyUI pads the channels and rescales the size to your checkpoint. Match the dropdown to what you're actually loading.

Sizes snap, on purpose

width/height default to 1024 and step in 16s on the widget, but the effective step follows the latent kind: 8 for SDXL/Anima/Boogu, 32 for Hunyuan, 16 for the rest. aspect_ratio (Free, 1:1, 16:9, 9:16, 4:3, 3:4, 21:9) only locks the step when it's Free; a fixed ratio anchors on the larger of your two numbers, computes the other side, then snaps it. Type 1000, get 1008 - VAE divisibility, not a bug. batch_size (default 1) is the batch dimension: four images, one prompt, four seeds.

One caveat that still applies: SDXL's trained buckets are 1024², 1152×896, 1216×832, 1344×768, 1536×640, and a 16:9 snap at a 1024 anchor gives 1024×576, outside them. Generate on-bucket and crop. On Z-Image, Flux 2 and Anima the discrete-bucket constraint has loosened a lot.

The prompt side

output_lang (中文/英文) sets the language of the preset sentence and the joiner that assembles the prompt. preset_mode is either 常规文生图 - your body text only, fully under your control - or 人物三视图, which prepends a character-sheet sentence asking for four panels: face close-up, then full-body front, side and back. That sentence lives in three_view_text, is editable, and only applies in three-view mode. It's the turnaround-sheet practice, the usual way to lock a character's angles and build LoRA datasets.

The panel is a row of category buttons (style, angle, subject, pose, outfit, props, lighting, background), each opening a chip list that assembles into your prompt. Your body text lives in internal_prompt, the same string the panel's preview box shows - edit either, the other follows. manager_settings stores the chips per node, so two of these don't interfere. Both fields are advanced, so they sit in the node's properties view rather than on its face.

A craft note, because a picklist isn't a prompt: on SDXL-lineage and Anima models you want booru tags, which is exactly the shape the chip output gives you. On Z-Image, Flux 2 and Qwen-Image the encoder is an LLM reading your prompt as an instruction, so use the chips as a skeleton and write a sentence around them.

Install

ComfyUI Manager, search XB_ToolBox, restart. Manually:

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

The README's "NO extra pip dependencies required!" is only true of the core nodes. The pack's requirements.txt pulls PyAV, opencv-python, easyocr, onnxruntime, pythonnet and more, and __init__.py imports the whole node set inside one try/except - so a single missing dependency doesn't cost you one node, it prints 🚨 [XB-BOX] FATAL ERROR: XB_ToolBox Loading Failed! and registers nothing. Install Manager's way, which runs requirements for you. Nothing to download for this node - it builds the latent itself.

Common issues

• Prompt box looks empty. Your text is in the advanced internal_prompt field. If the panel fails to render, restart and hard-refresh the browser tab; you can still edit that value from the node's properties.
• Half the menu is Chinese. The pack ships locales/zh and locales/en, but only some nodes are translated. Search "XB" in the add-node menu.
• My prompt starts with a paragraph. preset_mode is still on 人物三视图.
• The size changed after queuing. Wrong latent_kind: ComfyUI rescaled the latent to fit the loaded model.