{
  "article_slug": "visual-tokenization-rate",
  "figures": [
    {
      "id": "toy-rate-distortion",
      "kind": "line_chart",
      "title": "Capacity and fidelity in a four-image source",
      "x_label": "Decoder code states",
      "y_label": "Exact finite-source statistic",
      "x": [
        1,
        2,
        4
      ],
      "series": [
        {
          "name": "Maximum exact reconstruction probability",
          "y": [
            0.25,
            0.5,
            1
          ]
        },
        {
          "name": "Minimum normalized Hamming distortion",
          "y": [
            0.5,
            0.25,
            0
          ]
        }
      ],
      "caption": "Exhaustively enumerable author-constructed binary images. Not a learned-tokenizer benchmark or perceptual image metric."
    },
    {
      "id": "token-cost-scaling",
      "kind": "line_chart",
      "title": "Different quantities scale differently",
      "x_label": "Token positions L",
      "y_label": "Ratio to the L=256 configuration",
      "x": [
        32,
        64,
        128,
        256,
        512,
        1024
      ],
      "series": [
        {
          "name": "Fixed-width indices",
          "y": [
            0.125,
            0.25,
            0.5,
            1,
            2,
            4
          ]
        },
        {
          "name": "Dense-attention pairs",
          "y": [
            0.015625,
            0.0625,
            0.25,
            1,
            4,
            16
          ]
        }
      ],
      "caption": "Constructed counts with fixed alphabet K=1024. Ratios exclude entropy coding and all other network/system costs; they are not measured speedups."
    }
  ],
  "constructed_only": true
}