ComfyUI Node
MOSS-Image 提示词增强器 · Prompt Upsampler 🐑
A ComfyUI node in 🐑 MieNodes/🐑 Prompt Generator with 14 inputs and 3 outputs.
MOSS-Image 提示词增强器 · Prompt Upsampler 🐑
- llm_service_connector
- ref_image
- elements
- upsampled_prompt
- caption
- layout_json
◄modet2i - 文生图(只扩写描述)►
◄prompt►
◄canvas_width1024►
◄canvas_height1024►
◄seed0►
◄character_boxes►
◄keep_layout_verbatimtrue►
◄skip_enhancefalse►
◄temperature0.70►
◄max_tokens16384►
◄timeout300►
Category🐑 MieNodes/🐑 Prompt Generator
Inputs (14)
| Name | Type | Default | Description |
|---|---|---|---|
| llm_service_connector | LLMServiceConnector | — | |
| mode | COMBO | t2i - 文生图(只扩写描述) | 2 options: t2i - 文生图(只扩写描述), t2i_layout - 文生图+布局(生成Layout JSON) |
| prompt | STRING | 待扩写的简短请求，可中文；输出恒为英文描述，仅引号内待渲染文字保留原文。布局模式下还用于说明整张画面。 Short request to expand. Any language is fine -- the caption is always written in English, and only text the model is asked to render (inside double quotes) keeps its original language. | |
| canvas_width | INT | 102416–16384 | 画布宽度（像素，16 的倍数）。仅布局模式使用；必须与出图工作流的 width 一致。 Canvas width in pixels (a multiple of 16). Layout mode only; must match the width of the image workflow. |
| canvas_height | INT | 102416–16384 | 画布高度（像素，16 的倍数）。仅布局模式使用；必须与出图工作流的 height 一致。 Canvas height in pixels (a multiple of 16). Layout mode only; must match the height of the image workflow. |
| seed | INT | 00–18446744073709550000 | — |
| ref_imageopt | IMAGE | 只给 LLM 看的参考图（最多 4 张）。MOSS 本身不做图生图，图像模型不会收到这张图。 Reference image for the LLM only (up to 4). MOSS has no image-to-image mode -- the image model never sees this. | |
| elementsopt | ARRAY | Create Bounding Boxes 的 elements 输出。坐标是0-1000 网格的 [y_min, x_min, y_max, x_max]，本节点会自动换算成 MOSS 要求的绝对像素 xyxy，无需手写JSON。仅布局模式使用。 Elements output from Create Bounding Boxes. The boxes arrive on the official 0-1000 grid as [y_min, x_min, y_max, x_max]; the node converts them to the absolute-pixel xyxy that MOSS expects, so no JSON is needed. Layout mode only. | |
| character_boxesopt | STRING | 把哪些元素框标成角色（人物），填元素序号，如 "1, 3"。留空则全部按物体处理。仅布局模式使用。 Numbers of the element boxes to label as characters instead of objects, e.g. "1, 3". Layout mode only. | |
| keep_layout_verbatimopt | BOOLEAN | true | 勾选时以你画的框为准：官方文档明确指出提示词扩写会改写布局，勾选可保证框的像素不被模型改写，只取用模型的描述文字。取消勾选则完全采用模型规划的布局。 Keep the boxes you drew. The official docs warn that the prompt upsampler rewrites the layout, so this keeps your pixels and takes only the descriptions from the model. Turn it off to let the model plan the layout entirely. |
| skip_enhanceopt | BOOLEAN | false | 跳过扩写，直接用你的提示词（布局模式则直接把画好的框编译成 Layout）。官方文档把"关掉扩写以自己控制措辞"列为一项提示词技巧。 Skip the expansion and use your own wording. The official guide lists turning the upsampler off as a prompting technique. |
| temperatureopt | FLOAT | 0.700–2 | — |
| max_tokensopt | INT | 1638464–32768 | — |
| timeoutopt | COMBO | 300 | 4 options: 60, 120, 300, 600 |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| upsampled_prompt | STRING | — |
| caption | STRING | — |
| layout_json | STRING | — |