{
  "schema_version": 1,
  "sources": {
    "generation_table": {
      "path": "tab/tab_combine.tex",
      "sha256": "65a70b3bfccd4eda82e26664761acaf00fcb081f97cc3726c2b9db4ca0e15ba8"
    },
    "robotics_table": {
      "path": "tab/tab_robotic.tex",
      "sha256": "1bc554ea0a4e68ce8f335485c3ba4260770adf5744bdf7009398ee13e1f08612"
    },
    "experiment_protocol": {
      "path": "sec/05_experiment.tex",
      "sha256": "d979b5fd333d561c9ebf0969fa03947aacf0b946980597cf0bd7a30923170bd7"
    }
  },
  "mnist_ffhq": {
    "source": "tab/tab_combine.tex",
    "table_label": "tab:mnist_ffhq_comparison",
    "protocol": {
      "mnist": "Class-conditional generation in a learned 16-dimensional autoencoder latent space. EMD is average per-class squared 2-Wasserstein distance between test and generated latents. Accuracy uses a classifier trained on the MNIST training split.",
      "ffhq": "Six-class generation in frozen 512-dimensional ALAE latent space. Latent EMD and FID are reported; FID compares generated and real latents decoded by the same frozen ALAE decoder at 1024 x 1024."
    },
    "metrics": {
      "mnist_emd": {
        "label": "MNIST latent EMD",
        "direction": "lower"
      },
      "mnist_accuracy_percent": {
        "label": "MNIST accuracy",
        "unit": "percent",
        "direction": "higher"
      },
      "ffhq_emd": {
        "label": "FFHQ latent EMD",
        "direction": "lower"
      },
      "ffhq_fid": {
        "label": "FFHQ decoded-reference FID",
        "direction": "lower"
      }
    },
    "rows": [
      {
        "method": "Flow Matching",
        "nfe": 50,
        "mnist_emd": 37.2,
        "mnist_accuracy_percent": 100.0,
        "ffhq_emd": 198.4,
        "ffhq_fid": 77.4
      },
      {
        "method": "Mean Flow",
        "nfe": 1,
        "mnist_emd": 68.6,
        "mnist_accuracy_percent": 96.7,
        "ffhq_emd": 258.7,
        "ffhq_fid": 131.2
      },
      {
        "method": "Mean Flow",
        "nfe": 2,
        "mnist_emd": 48.1,
        "mnist_accuracy_percent": 99.1,
        "ffhq_emd": 211.7,
        "ffhq_fid": 85.4
      },
      {
        "method": "Mean Flow",
        "nfe": 5,
        "mnist_emd": 37.5,
        "mnist_accuracy_percent": 100.0,
        "ffhq_emd": 204.2,
        "ffhq_fid": 82.1
      },
      {
        "method": "Drift Model",
        "nfe": 1,
        "mnist_emd": 37.3,
        "mnist_accuracy_percent": 100.0,
        "ffhq_emd": 225.5,
        "ffhq_fid": 116.2
      },
      {
        "method": "DFM",
        "nfe": 1,
        "mnist_emd": 37.3,
        "mnist_accuracy_percent": 100.0,
        "ffhq_emd": 225.5,
        "ffhq_fid": 116.2
      },
      {
        "method": "DFM",
        "nfe": 2,
        "mnist_emd": 37.2,
        "mnist_accuracy_percent": 100.0,
        "ffhq_emd": 215.7,
        "ffhq_fid": 80.4
      },
      {
        "method": "DFM",
        "nfe": 5,
        "mnist_emd": 37.2,
        "mnist_accuracy_percent": 100.0,
        "ffhq_emd": 196.2,
        "ffhq_fid": 75.9
      }
    ]
  },
  "imagenet": {
    "source": "tab/tab_combine.tex",
    "table_label": "tab:image256",
    "protocol": "ImageNet-1k 256 x 256 synthesis using a pretrained VAE tokenizer with 32 x 32 x 4 latents and a latent-MAE feature-space DFM objective. FID and IS are evaluated on 50,000 randomly generated images, following the Drift Model protocol.",
    "model": "DFM, L/2",
    "parameters_millions": 463,
    "metrics": {
      "fid": {
        "label": "FID",
        "direction": "lower"
      },
      "inception_score": {
        "label": "IS",
        "direction": "higher"
      }
    },
    "rows": [
      {
        "nfe": 1,
        "fid": 1.52,
        "inception_score": 259.4
      },
      {
        "nfe": 2,
        "fid": 1.45,
        "inception_score": 268.3
      },
      {
        "nfe": 5,
        "fid": 1.34,
        "inception_score": 279.8
      },
      {
        "nfe": 10,
        "fid": 1.31,
        "inception_score": 287.4
      }
    ]
  },
  "robotics": {
    "source": "tab/tab_robotic.tex",
    "table_label": "tab:robotics_control",
    "protocol": "The core policy is replaced with DFM under the Diffusion Policy and Drift Policy experimental setting. DFM loss uses raw control representations. Reported success rates average the last 10 checkpoints.",
    "method": "Drift Flow Matching Policy",
    "metric": {
      "label": "Success rate",
      "unit": "fraction",
      "direction": "higher"
    },
    "nfe": [
      1,
      2,
      5,
      10
    ],
    "series": [
      {
        "task": "ToolHang",
        "setting": "State",
        "success_rate": [
          0.41,
          0.77,
          0.81,
          0.86
        ]
      },
      {
        "task": "PushT",
        "setting": "Visual",
        "success_rate": [
          0.86,
          0.87,
          0.89,
          0.9
        ]
      }
    ]
  }
}