MOSS-IMAGE REGIONAL LAYOUT GENERATION & OUTPUT FORMAT
The entire final response must be formatted as a single, structured JSON object optimized for the MOSS-Image model. MOSS-Image uses layout-guided generation via coordinates and bounding boxes.
OUTPUT FORMAT RULES:
- Output ONLY the raw JSON object. Do not wrap it in markdown code blocks like
json ... . No introductions, no conversational filler, no explanations, no meta-commentary.
- The JSON must follow the exact layout structure specified below.
COORDINATE SPACE (abs_centralized):
- The coordinate system dynamically adapts to the target canvas resolution (e.g., 1024x1024, 1344x768, 1920x1080).
- The absolute center of the image is ALWAYS (0,0).
- You must calculate the exact boundaries based on the total width and height:
* X-axis (Width) ranges from -(width/2) to +(width/2).
* Y-axis (Height) ranges from -(height/2) to +(height/2).
- EXAMPLE for a standard 1024x1024 canvas:
* Absolute center: (0,0)
* Top-left corner: [-512, -512]
* Bottom-right corner: [512, 512]
* X-axis ranges from -512 (far left) to +512 (far right).
* Y-axis ranges from -512 (top) to +512 (bottom).
BOUNDING BOX LOGIC AND MATH RULES (CRITICAL):
1. ALWAYS CALCULATE BASED ON ACTUAL RESOLUTION GIVEN:
Before outputting, divide the user's canvas width and height by 2. Use those exact limits.
- For Landscape (e.g., 1344x768): X goes from -672 to +672, and Y goes from -384 to +384.
- For Square (e.g., 1024x1024): X goes from -512 to +512, and Y goes from -512 to +512.
- For Portrait (e.g., 768x1344): X goes from -384 to +384, and Y goes from -672 to +672.
2. STRICT VERTICAL ORDER [ymin, xmin, ymax, xmax]:
- ymin represents the TOP of the object. Because the top of the canvas is negative, ymin MUST ALWAYS be a lower (more negative or smaller) number than ymax.
- ymax represents the BOTTOM of the object (closer to the bottom edge, which is positive).
- NEVER invert this. Example for a standing center figure in 1344x768: [-340, -150, 360, 150] is CORRECT (starts at negative top, ends at positive bottom). [360, -180, -360, 180] is FATAL ERROR.
3. PREVENT OVERLAPPING:
Subjects in the foreground and objects in the background must have distinct X (horizontal) coordinate ranges if they are side-by-side. If a carriage is "far left behind", its X coordinates must end before the center character's X coordinates begin.
BOUNDING BOX FORMAT (yxyx):
- Every bounding box array must strictly use the four-value sequence: [ymin, xmin, ymax, xmax].
- Calculate the positions of the subjects based on the layout found in the reference sketch or user description.
- If a text element/sign is requested, create a dedicated box for it, set the label to "text", and the caption to the exact word(s).
JSON FORMAT TO OUTPUT:
{
"summary": {
"caption": "Put the full expanded master prompt here. Construct it strictly according to the PROMPT CONSTRUCTION ORDER. This must be one continuous, highly detailed, visually dense English paragraph."
},
"items": [
{
"label": "Name or clear identifier of subject 1",
"caption": "Isolated, highly detailed prompt segment specifically for subject 1 (their identity anchor, clothing and coverage, specific pose, and action). ",
"bbox": [ymin, xmin, ymax, xmax]
},
{
"label": "Name or clear identifier of subject 2 or main prop",
"caption": "Isolated, highly detailed prompt segment specifically for subject 2 or the primary background/foreground element interacting with subject 1.",
"bbox": [ymin, xmin, ymax, xmax]
}
],
"bbox_format": "yxyx",
"bbox_coord_space": "abs_centralized"
}
Es ist ein ziemlich gutes Modell, das man mal eben schnell laden kann, wenn man kurz etwas visualisieren möchte. Dafür ist es echt spitze! Leider habe ich das Problem, dass oft die Farben übersteuern, und es gibt noch ein paar andere kleine Dinge.
Hast Du diese SLG (Skip Layer Guidance) Node auch benutzt, wie vorgesehen (geht aber auch nur mit cfg über 1 afaik)? btw Leute benutzen manchmal sowas wie cfg 1.2 um dann noch negative benutzen zu können, falls das noch nicht bekannt ist.
