grid-image-generator

Image Prompt Generation Guide

This SKILL corresponds to grid_prompt_generator Agent, supporting the generation of three types of image prompts:

1. Character Image Prompts — Character appearance and temperament
2. Scene Image Prompts — Scene atmosphere and lighting
3. Grid Image Prompts — Multi-shot grid collage

For detailed templates, see the reference/ directory.

---

Character Image Prompts

Reference:reference/character-prompt.md

Template Structure

[appearance], [personality/temperament], [role], [cinematic portrait], [high quality], [consistent art style], [no text, no watermark]

Generation Rules

• Centered on appearance(appearance description)
• personality determines the temperament tone (reserved/bold/mysterious, etc.)
• role determines the clothing and prop style
• Must include cinematic portrait + consistent art style
• Avoid text, signatures, and watermarks

---

Scene Image Prompts

Reference:reference/scene-prompt.md

Template Structure

[location], [time period], [lighting atmosphere], [scene description], [cinematic scene], [high quality], [consistent art style], [no text, no watermark]

Generation Rules

• Based on location(location)
• time determines the lighting and color tone (day/night/dusk)
• Scene atmosphere words: atmospheric, moody, warm, cold, etc.
• Must include cinematic scene + consistent art style
• Avoid text, signatures, and watermarks

---

Grid Image Prompts

Reference:reference/shot-prompt.md

Three Modes

First Frame Mode (first_frame)

Each cell = the starting frame of one shot, but must strictly generate the user-specified rows x cols total number of cells.

[rows x cols grid layout], exactly [rows*cols] visible panels, consistent art style, [style description],
格1: [shot 1 opening scene],
格2: [shot 2 opening scene],
格3: [shot 3 opening scene],
...
格N: [opening scene],
high quality, cinematic lighting, no merged panels, no missing panels, no text, no watermark

First and Last Frame Mode (first_last)

Maintain the rhythm of first and last frames, but must still strictly generate the user-specified rows x cols total number of cells, and must not secretly change it to Nx2。

[rows x cols grid layout], exactly [rows*cols] visible panels, consistent art style, [style description],
格1: [opening beat],
格2: [closing beat],
格3: [opening beat],
格4: [closing beat],
...
high quality, cinematic, continuous motion implied, no merged panels, no missing panels, no text

Multi-Reference Mode (multi_ref)

All cells are different angle/composition references of the same shot, but must still strictly generate the user-specified rows x cols total number of cells.

[rows x cols grid layout], exactly [rows*cols] visible panels, same scene different angles, [style description],
[main scene description],
格1: wide shot establishing,
格2: medium shot character focus,
格3: close-up detail,
格4: dramatic angle,
...
consistent lighting and color palette, no merged panels, no missing panels, no text

General Rules

1. Prompts useEnglish
2. Must explicitly write the user-specified rows x cols grid layout
3. Must include consistent art style Keep the style consistent
4. Must explicitly require exactly N visible panels
5. Must explicitly require no merged panels, no missing panels
6. Avoid descriptions of dividing lines between cells
7. Size suggestion: 960x540 per cell, total image = 960×cols × 540×rows
8. When a reference image mapping exists, uniformly use 图片1/图片2/... to refer to the reference image, and do not mix it with 格1/格2/... .