world-labs-multi-image（世界实验室多图像）

World Labs Multi-Image Input

Generate 3D worlds from multiple reference images with precise directional control or automatic layout.

Quick Reference

Credits

Direction Control Mode

Use when you want explicit control over image placement. Non-overlapping images are preferred—the model creatively fills spaces between views.

Azimuth Angles

Any value 0-360 is supported for non-cardinal directions.

Visual Reference

              0° (Front)
                  ↑
                  |
    270° (Left) ←─┼─→ 90° (Right)
                  |
                  ↓
            180° (Back)

API Usage (Direction Control)

{
  "model": "Marble 0.1-plus",
  "world_prompt": {
    "type": "multi-image",
    "multi_image_prompt": [
      {
        "azimuth": 0,
        "content": {
          "source": "media_asset",
          "media_asset_id": "front_image_id"
        }
      },
      {
        "azimuth": 90,
        "content": {
          "source": "media_asset",
          "media_asset_id": "right_image_id"
        }
      },
      {
        "azimuth": 180,
        "content": {
          "source": "media_asset",
          "media_asset_id": "back_image_id"
        }
      },
      {
        "azimuth": 270,
        "content": {
          "source": "media_asset",
          "media_asset_id": "left_image_id"
        }
      }
    ],
    "text_prompt": "A grand Victorian mansion interior"
  }
}

Using Public URLs

{
  "model": "Marble 0.1-plus",
  "world_prompt": {
    "type": "multi-image",
    "multi_image_prompt": [
      {
        "azimuth": 0,
        "content": {
          "source": "uri",
          "uri": "https://example.com/front.jpg"
        }
      },
      {
        "azimuth": 180,
        "content": {
          "source": "uri",
          "uri": "https://example.com/back.jpg"
        }
      }
    ],
    "text_prompt": "A cozy living room"
  }
}

Best Practices for Direction Control

✅ Non-overlapping images: Model creatively fills gaps between views ✅ Consistent style: All images should match aesthetically ✅ Same lighting: Consistent light direction across all images ✅ Complementary angles: Choose angles that tell a complete story

Auto Layout Mode

Use when you have multiple overlapping images from the same space. The model automatically positions images.

Requirements for Auto Layout

API Usage (Auto Layout)

Auto Layout is primarily a UI feature. In API, omit azimuth values:

{
  "model": "Marble 0.1-plus",
  "world_prompt": {
    "type": "multi-image",
    "multi_image_prompt": [
      {
        "content": {
          "source": "media_asset",
          "media_asset_id": "image_1_id"
        }
      },
      {
        "content": {
          "source": "media_asset",
          "media_asset_id": "image_2_id"
        }
      },
      {
        "content": {
          "source": "media_asset",
          "media_asset_id": "image_3_id"
        }
      }
    ],
    "text_prompt": "A mystical forest clearing"
  }
}

Python Example

import requests

def generate_world_multi_image(
    api_key: str,
    images: list[dict],  # [{"path": "...", "azimuth": 0}, ...] or [{"path": "..."}]
    prompt: str = None
):
    base_url = "https://api.worldlabs.ai/marble/v1"
    headers = {"WLT-Api-Key": api_key, "Content-Type": "application/json"}

    multi_image_prompt = []

    for img in images:
        # Get extension
        ext = img["path"].lower().split('.')[-1]
        if ext == "jpeg":
            ext = "jpg"

        # Prepare upload
        prep = requests.post(
            f"{base_url}/media-assets:prepare_upload",
            headers=headers,
            json={"file_name": img["path"].split('/')[-1], "kind": "image", "extension": ext}
        ).json()

        media_asset_id = prep["media_asset"]["media_asset_id"]
        upload_url = prep["upload_info"]["upload_url"]

        # Upload file
        content_types = {"jpg": "image/jpeg", "png": "image/png", "webp": "image/webp"}
        with open(img["path"], 'rb') as f:
            requests.put(upload_url, headers={"Content-Type": content_types.get(ext, "image/jpeg")}, data=f.read())

        # Build image reference
        image_entry = {
            "content": {
                "source": "media_asset",
                "media_asset_id": media_asset_id
            }
        }
        if "azimuth" in img:
            image_entry["azimuth"] = img["azimuth"]

        multi_image_prompt.append(image_entry)

    # Build world_prompt
    world_prompt = {
        "type": "multi-image",
        "multi_image_prompt": multi_image_prompt
    }
    if prompt:
        world_prompt["text_prompt"] = prompt

    # Generate
    response = requests.post(
        f"{base_url}/worlds:generate",
        headers=headers,
        json={"model": "Marble 0.1-plus", "world_prompt": world_prompt}
    )

    return response.json()["operation_id"]

# Usage with direction control
operation_id = generate_world_multi_image(
    "your_api_key",
    [
        {"path": "front_view.jpg", "azimuth": 0},
        {"path": "right_view.jpg", "azimuth": 90},
        {"path": "back_view.jpg", "azimuth": 180},
        {"path": "left_view.jpg", "azimuth": 270},
    ],
    prompt="A cozy cabin interior"
)

# Usage with auto layout (no azimuth)
operation_id = generate_world_multi_image(
    "your_api_key",
    [{"path": f"ref_{i}.jpg"} for i in range(6)],
    prompt="An enchanted garden"
)

Common Use Cases

Interior Room (4 corners)

{
  "world_prompt": {
    "type": "multi-image",
    "multi_image_prompt": [
      {
        "azimuth": 45,
        "content": { "source": "media_asset", "media_asset_id": "corner_1" }
      },
      {
        "azimuth": 135,
        "content": { "source": "media_asset", "media_asset_id": "corner_2" }
      },
      {
        "azimuth": 225,
        "content": { "source": "media_asset", "media_asset_id": "corner_3" }
      },
      {
        "azimuth": 315,
        "content": { "source": "media_asset", "media_asset_id": "corner_4" }
      }
    ]
  }
}

Street Scene (looking both ways)

{
  "world_prompt": {
    "type": "multi-image",
    "multi_image_prompt": [
      {
        "azimuth": 0,
        "content": { "source": "media_asset", "media_asset_id": "street_north" }
      },
      {
        "azimuth": 180,
        "content": { "source": "media_asset", "media_asset_id": "street_south" }
      }
    ]
  }
}

Comparison: Direction Control vs Auto Layout

Troubleshooting

Related Skills

• world-labs-api - API integration details
• world-labs-image-prompt - Single image input
• world-labs-pano-video - Panorama and video input (most control)