MOSS-Image: make a poster layout you can check

1. Set the finished output before generating

Make one square image containing the two exact text strings and one orange desk lamp. The title belongs above the picture; the time belongs below it. Keep the cream background and dark green lettering. You should be able to read both lines when the image is shown at 320 pixels wide.

Keep a folder containing poster.json, your saved image and a short review record. A successful result has correct wording, sensible spacing and an obvious reading order. A beautiful picture with a misspelled title fails this task. Do not add real event details until the fictional version passes.

2. Save the complete layout prompt

Copy the following JSON into a plain UTF-8 file named poster.json. It is an original exercise fixture. A bounding box describes the requested rectangular area for an item. For this 1024-pixel square, coordinates run from -512 to 512 around the center; the order is top, left, bottom, right.

The three boxes deliberately leave gaps. The title has a wide band at the top, the illustration occupies the middle, and the time sits near the bottom. These are requested regions, so the generated image still needs a visual check.

{
  "summary": {
    "meta": [
      "poster"
    ],
    "canvas": "single",
    "caption": "A square poster for a fictional repair cafe. Warm cream paper, dark green lettering, and a simple orange desk-lamp illustration. Use only the two supplied text strings. Keep generous empty space around each element."
  },
  "items": [
    {
      "label": "text",
      "caption": "Large dark green uppercase title centered near the top.",
      "text": "REPAIR CAFE",
      "bbox": [
        -420,
        -410,
        -270,
        410
      ]
    },
    {
      "label": "object",
      "caption": "One simple orange desk lamp with a wide base, drawn as a flat illustration.",
      "bbox": [
        -200,
        -220,
        190,
        220
      ]
    },
    {
      "label": "text",
      "caption": "Smaller dark green time label centered below the illustration.",
      "text": "SATURDAY 10-12",
      "bbox": [
        280,
        -350,
        390,
        350
      ]
    }
  ],
  "bbox_coord_space": "abs_centralized",
  "bbox_format": "yxyx"
}

3. Check the numbers before judging the model

If you have Python, save the next block as check_layout.py in the same folder. It uses only the standard library and does not contact a service. The helper checks this fixture’s geometry and required text; it is not a validator for every feature of the model’s input format.

Open a terminal in that folder and run the command below. Compare the printed corners with the expected output. The title starts 102 pixels from the left and 92 pixels from the top. The time starts 162 pixels from the left and 792 pixels from the top. These positions give you a simple reference when you inspect the image.

import json
import sys
from pathlib import Path

def check(layout, width=1024, height=1024):
    if width <= 0 or height <= 0 or width % 16 or height % 16:
        raise ValueError("Dimensions must be positive multiples of 16")
    if layout.get("bbox_format") != "yxyx":
        raise ValueError("Use yxyx coordinate order")
    if layout.get("bbox_coord_space") != "abs_centralized":
        raise ValueError("Use centered pixel coordinates")
    items = layout.get("items")
    if not isinstance(items, list) or not items:
        raise ValueError("Provide at least one item")
    for i, item in enumerate(items, 1):
        box = item.get("bbox")
        if not isinstance(box, list) or len(box) != 4:
            raise ValueError(f"Item {i}: expected four coordinates")
        if any(type(v) not in (int, float) for v in box):
            raise ValueError(f"Item {i}: coordinates must be numbers")
        y0, x0, y1, x1 = box
        if not (-height/2 <= y0 < y1 <= height/2
                and -width/2 <= x0 < x1 <= width/2):
            raise ValueError(f"Item {i}: reversed or out-of-bounds box")
        if item.get("label") == "text" and not item.get("text", "").strip():
            raise ValueError(f"Item {i}: missing text")
        print(f"Item {i}: top-left=({x0+width/2:g}, {y0+height/2:g}), "
              f"bottom-right=({x1+width/2:g}, {y1+height/2:g})")
    print(f"PASS: {len(items)} boxes fit a {width}x{height} canvas")

if __name__ == "__main__":
    try:
        check(json.loads(Path(sys.argv[1]).read_text(encoding="utf-8")))
    except (IndexError, OSError, ValueError, KeyError, TypeError) as error:
        raise SystemExit(f"FAIL: {error}")

python3 check_layout.py poster.json

Item 1: top-left=(102, 92), bottom-right=(922, 242)
Item 2: top-left=(292, 312), bottom-right=(732, 702)
Item 3: top-left=(162, 792), bottom-right=(862, 902)
PASS: 3 boxes fit a 1024x1024 canvas

The supplied fixture passed this check. Deliberately reversed, oversized and missing-text variants failed. If Python is unavailable, you can still use the prompt: compare the four numbers of each box with the coordinate convention above and record that the automatic check was skipped. A passing geometry check says nothing about the quality of a generated image.

4. Try one image in the public demo

Open the MOSS-Image public demo from Source documentation below. Select MOSS-Image-Base. Paste the complete JSON into Prompt. Set width and height to 1024, image count to 1, seed to 42 and steps to 50. Leave other settings at their displayed defaults and record them in your notes. This guide follows the Base layout documentation; it does not assume that another checkpoint will follow the same composition.

Press Generate only when the interface offers access you are comfortable using. A queue, account requirement or resource limit can prevent a run. The public page did not establish a guaranteed free allowance or completion time. Stop if it requires an unexpected purchase; you can keep the prepared fixture for a later attempt.

When an image is returned, save it using the result’s download control if available. If saving is unavailable, record that export remains unverified. Preserve the exact prompt and settings with the image. No model inference was performed when preparing this article; these are documentation-based instructions for your own trial.

5. Accept or reject the visible result

Check wording first, then placement, then appearance. Read each character in REPAIR CAFE and SATURDAY 10-12. Reject extra words, invented logos and an incorrect time. View the image at 320 pixels wide to check whether the text still works on a phone.

The lamp should sit between the two text bands without covering them. Use the expected corner coordinates as a reference for the composition, allowing for visual margins inside each box. If the text is right but the illustration crosses a band, record a placement failure. If the composition is right but letters are malformed, record a lettering failure. Keeping those outcomes separate makes the next change clearer.

Input file: poster.json
Model: MOSS-Image-Base
Canvas: 1024x1024
Seed: 42
Steps: 50
Other displayed settings: record their actual values
Exact title: PASS / FAIL
Exact time: PASS / FAIL
No additional text or logos: PASS / FAIL
Title above lamp, time below lamp: PASS / FAIL
Both lines readable at 320px wide: PASS / FAIL
Saved output filename: record the actual filename
Decision: ACCEPT only when every visual check passes
Observed failure and next edit: describe one issue

6. Make one measurable revision

For a second attempt, change only the lamp box to [-200,-300,190,140]. This moves the requested region 80 pixels left while preserving its size. Keep the canvas, seed, wording and other settings unchanged. Save the changed prompt as poster-left.json and rerun the coordinate helper.

Compare the two images side by side. The lamp should move left while both text lines remain readable and correctly spelled. A fixed seed helps keep the comparison controlled, but it does not guarantee identical pixels across a changed prompt, model build or runtime. If the wording also changes, record that as an unintended revision.

Keep both versions even if the second looks worse. The useful output is a traceable design decision: what changed in the prompt, what changed in the image, and whether the result passed. The editable artifact here is the composition JSON; a raster image does not provide separate vector text and illustration layers.

When something fails

For an actual event poster, finish with a human check of the venue, date, access information and permission to use any imagery. Those details are intentionally absent from this fictional exercise. Model weights being public does not remove inference costs: local use requires suitable hardware, and hosted access can have limits. Choose a deployment path only after the small composition test is useful to you.

Guide history

Revisions preserve the original instructions. A source recheck alone does not create a new revision.

Revision 1 · New original poster-layout exercise with complete JSON, a tested coordinate helper, exact browser steps, visual acceptance checks and a controlled revision. Image inference remains unexecuted.