{
  "format": "companion-forge-benchmarks/v1",
  "hardware": {
    "flavor": "l4x1",
    "gpu": "NVIDIA L4",
    "compute_capability": "8.9",
    "usable_vram_gib": 22.03
  },
  "dense_modules": {
    "dinov2": {
      "ort_cuda_mean_ms": 98.988,
      "trt_mean_ms": 20.501,
      "speedup_vs_ort": 4.8285,
      "cosine": 0.9999338
    },
    "dsine": {
      "pytorch_mean_ms": 88.918,
      "ort_cuda_mean_ms": 201.281,
      "trt_mean_ms": 24.581,
      "speedup_vs_pytorch": 3.617,
      "cosine_vs_pytorch": 0.9996713
    },
    "flux2_transformer_512": {
      "pytorch_mean_ms": 236.838,
      "trt_mean_ms": 208.095,
      "speedup_vs_pytorch": 1.1381,
      "cosine": 0.999836
    }
  },
  "end_to_end": {
    "text_to_reference_trt512": {
      "seconds": 103.499,
      "peak_device_vram_gib": 10.062,
      "runtime": "trt512",
      "steps": 4,
      "resolution": 512
    },
    "image_to_rigged_fast": {
      "seconds": 115.521,
      "peak_device_vram_gib": 11.406,
      "runtime": "hybrid-staged-dino-trt+dsine-trt",
      "sampling_steps": "10+10",
      "joints_test_asset": 17,
      "vertices_test_asset": 5964,
      "faces_test_asset": 11879
    }
  }
}