eenvil/stage3-relighting-dataset

Stage 3 Relighting Dataset

This is a self-contained Parquet version of the Stage 3 relighting dataset. Each row represents one scene. Image columns are embedded as Hugging Face Image features, so the dataset can be loaded without separately downloading a large tree of individual image files.

Stage 3 Task

The model learns to map optimized physically based renders back toward the original real photograph:

inputs  -> neural renderer -> prediction
target  -> reference

The canonical best current input set is:

composite_optimized
render_optimized
hybrid_albedo
roughness
metallic
glass_mask_soft

This corresponds to a 12-channel tensor after preprocessing:

composite RGB      3
render RGB         3
hybrid_albedo RGB  3
roughness gray     1
metallic gray      1
glass_mask gray    1

Splits

{
  "train": {
    "rows": 18510,
    "skipped": 0,
    "shards": 90,
    "image_gib": 89.67,
    "optional_present": {
      "hybrid_albedo": 18510,
      "roughness": 18510,
      "metallic": 18510,
      "glass_mask_soft": 18510,
      "normal": 18510,
      "shading": 18510
    }
  },
  "validation": {
    "rows": 1028,
    "skipped": 0,
    "shards": 5,
    "image_gib": 5.01,
    "optional_present": {
      "hybrid_albedo": 1028,
      "roughness": 1028,
      "metallic": 1028,
      "glass_mask_soft": 1028,
      "normal": 1028,
      "shading": 1028
    }
  },
  "test": {
    "rows": 1029,
    "skipped": 0,
    "shards": 5,
    "image_gib": 4.08,
    "optional_present": {
      "hybrid_albedo": 0,
      "roughness": 1029,
      "metallic": 1029,
      "glass_mask_soft": 0,
      "normal": 1029,
      "shading": 1029
    }
  }
}

The test split currently uses MV-inverse buffers only, so hybrid_albedo and glass_mask_soft are null there unless glass-aware buffers are generated for those scenes later.

Loading

from datasets import load_dataset

ds = load_dataset("eenvil/stage3-relighting-dataset")
train = ds["train"]

Training code should convert null optional image columns to zero tensors.