{
  "scope": "Single reference geometry before the VL processor not model tensors or inference",
  "source": {
    "width": 1200,
    "height": 800
  },
  "mode": "long-edge",
  "requested": {
    "longEdge": 1024,
    "targetW": 768,
    "targetH": 768,
    "vlCap": 384
  },
  "outputAndVae": {
    "width": 1024,
    "height": 672
  },
  "beforeVlProcessor": {
    "width": 384,
    "height": 256
  },
  "relativeAspectChange": 0.015873015873015817,
  "vaeResize": "Direct resize to output dimensions without a center crop",
  "visionResize": "Long edge cap without upscaling before the downstream VL processor"
}